<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: FX-LgLL</title>
    <description>The latest articles on DEV Community by FX-LgLL (@montydrief).</description>
    <link>https://dev.to/montydrief</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4093945%2F96c0f467-7f7c-4f5c-a971-041ca4ade958.jpeg</url>
      <title>DEV Community: FX-LgLL</title>
      <link>https://dev.to/montydrief</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/montydrief"/>
    <language>en</language>
    <item>
      <title>100 million tokens in 4 days: what developers built with an open 4.5B decision model</title>
      <dc:creator>FX-LgLL</dc:creator>
      <pubDate>Mon, 09 Nov 2026 14:23:55 +0000</pubDate>
      <link>https://dev.to/montydrief/100-million-tokens-in-4-days-what-developers-built-with-an-open-45b-decision-model-5b7e</link>
      <guid>https://dev.to/montydrief/100-million-tokens-in-4-days-what-developers-built-with-an-open-45b-decision-model-5b7e</guid>
      <description>&lt;p&gt;Four days ago we released &lt;strong&gt;THX-02&lt;/strong&gt;, an open 4.5B decision model. You send a state (text, JSON or images) and typed questions; it returns a calibrated probability for every option you give it, and it writes a short thought first only when a question needs reasoning.&lt;/p&gt;

&lt;p&gt;In those four days, more than &lt;strong&gt;100,000,000 tokens&lt;/strong&gt; went through THX-02 on our API. THX-02 also reached &lt;strong&gt;#1 in Zero-Shot Classification on Hugging Face&lt;/strong&gt;. Here is what people built, with no names, just the shapes of the problems.&lt;/p&gt;

&lt;h2&gt;
  
  
  Developer tools and AI agents
&lt;/h2&gt;

&lt;p&gt;The biggest group by far.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Pull-request review in CI.&lt;/strong&gt; For every diff: does it expose a secret? For the PR description: does it claim something the code doesn't do, include unrelated changes, or contradict its tests?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Guardrails for coding agents.&lt;/strong&gt; Does this command violate the policy, and did the user actually ask for it? Is a "done" claim backed by evidence? Is this session's learning worth saving to memory?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Routing.&lt;/strong&gt; Which tool to call first, which use case a request belongs to, when to hand off to a human.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Decision models fit here because the answer must be one of a fixed set, fast, and come with a confidence you can threshold.&lt;/p&gt;

&lt;h2&gt;
  
  
  Documents, law and the public sector
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Italian fact-checking: who speaks a quote, and whether a claim is supported by the passage.&lt;/li&gt;
&lt;li&gt;Italian long-document Q&amp;amp;A: does any line answer this rule, and where.&lt;/li&gt;
&lt;li&gt;Italian land and permits: landscape constraints and planning rules for solar.&lt;/li&gt;
&lt;li&gt;French public procurement: would a company with this profile bid on this tender.&lt;/li&gt;
&lt;li&gt;Indian GST bookkeeping: which ledger, asset or expense, and whether input tax credit is blocked.&lt;/li&gt;
&lt;li&gt;Job postings: 15 fields per posting, from region and seniority to visa sponsorship.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Marketing, support and vision
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Landing-page copy and UX screen audits, mostly with thinking on.&lt;/li&gt;
&lt;li&gt;Support ticket triage in English, Chinese and Azerbaijani.&lt;/li&gt;
&lt;li&gt;Receipts, tables and charts; video frames for sponsored-product visibility.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  And the unexpected
&lt;/h2&gt;

&lt;p&gt;An &lt;strong&gt;AI techno DJ&lt;/strong&gt; asking which color, melodic and rhythm layer to bring in for the next 16 bars. A game studio choosing which recorded voice line replaces the hero's. An agent walking through a maze.&lt;/p&gt;

&lt;h2&gt;
  
  
  What we learned
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Decision models shine as checkers inside other systems&lt;/strong&gt;: review, guardrails, moderation, routing. That's where confidence was highest (0.93 to 0.98 on average).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;People bring long, real inputs&lt;/strong&gt;: diffs, legal documents, transcripts. Most tokens came from texts of several thousand tokens.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;They are honest about what they can't do.&lt;/strong&gt; Numeric estimation (calories) and spatial puzzles got low confidence, which is the right behavior for a calibrated model, and a clear target for the next version.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;thx02
thx02 &lt;span class="s2"&gt;"My order arrived broken, please send a new one"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--choice&lt;/span&gt; &lt;span class="s2"&gt;"action=refund:money back,replace:send a new item,ignore:no action"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--yesno&lt;/span&gt; &lt;span class="s2"&gt;"Does this ticket need a reply?"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ul&gt;
&lt;li&gt;Weights (Apache-2.0): &lt;a href="https://huggingface.co/HAL-X-AI/THX-02" rel="noopener noreferrer"&gt;https://huggingface.co/HAL-X-AI/THX-02&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Docs and free API: &lt;a href="https://api.hal-x.ai/docs/thx-02" rel="noopener noreferrer"&gt;https://api.hal-x.ai/docs/thx-02&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;We believe AI should be built for everyone and with everyone. Thank you to everyone who integrated THX-02 in its first week.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;HAL-X, Baku&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>programming</category>
      <category>opensource</category>
      <category>automation</category>
    </item>
    <item>
      <title>How we made a 4.5B model think only when it changes the answer</title>
      <dc:creator>FX-LgLL</dc:creator>
      <pubDate>Fri, 06 Nov 2026 15:00:13 +0000</pubDate>
      <link>https://dev.to/montydrief/how-we-made-a-45b-model-think-only-when-it-changes-the-answer-36o4</link>
      <guid>https://dev.to/montydrief/how-we-made-a-45b-model-think-only-when-it-changes-the-answer-36o4</guid>
      <description>&lt;h1&gt;
  
  
  How we made a 4.5B model think only when it changes the answer
&lt;/h1&gt;

&lt;p&gt;There is a trap in building a decision model, and it is not the one you expect.&lt;/p&gt;

&lt;p&gt;The obvious approach is: give the model a chain-of-thought budget and let it reason before answering. It works. Accuracy goes up on the hard cases. And then you ship it, and every trivial request — "which of these four teams owns this ticket" — now costs you a thousand tokens of deliberation that changes nothing.&lt;/p&gt;

&lt;p&gt;Worse, on the cases where the model was already right, the reasoning can talk it out of the correct answer. A model that reasons itself from 96% correct into 51% wrong is not smarter. It is just slower and less reliable.&lt;/p&gt;

&lt;p&gt;THX-02 is a 4.5B decision model we released under Apache-2.0, with a free API. This post is about the specific mechanism we built to avoid that trap: a gate that decides whether thinking is worth doing, and a mixing step that makes a bad thought unable to overturn a confident answer.&lt;/p&gt;

&lt;p&gt;The numbers, up front, so you know what the mechanism buys: on math word problems the thought adds about 25 points. On natural language inference, code tracing and graduate-level science it adds 4 to 6. Median latency with the gate active stays around 80 ms. Direct answering alone scores 63.83 on Decision Index 0.3; with the gate, 65.06.&lt;/p&gt;




&lt;h2&gt;
  
  
  The shape of the problem
&lt;/h2&gt;

&lt;p&gt;A decision model is not a chat model with a different prompt. The interface is different in a way that changes everything downstream:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;thx02&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;Client&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;choice&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;noul&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;score&lt;/span&gt;

&lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;Client&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;

&lt;span class="n"&gt;result&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;decide&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;state&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Ticket: customer says the mobile app logs them out every few minutes &lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
          &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;after the latest update. Android 14, app version 4.2.1.&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;questions&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;
        &lt;span class="nf"&gt;choice&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;team&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;billing&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;support&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;shipping&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;technical&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;security&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]),&lt;/span&gt;
        &lt;span class="nf"&gt;noul&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;escalate&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Should this go to a senior engineer?&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt;
        &lt;span class="nf"&gt;score&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;severity&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;How severe is this for the customer?&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;5&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt;
    &lt;span class="p"&gt;],&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Every answer is a full distribution over the options you supplied. That is the contract. The model cannot emit a label that is not on your list, because there is no generation step — it scores each option with a marker token and normalizes.&lt;/p&gt;

&lt;p&gt;That constraint is what makes the gate possible in the first place. If the model were free-generating, "did thinking help?" would be a question about text similarity. Here it is a question about probability mass on a fixed set, which is measurable.&lt;/p&gt;




&lt;h2&gt;
  
  
  The gate
&lt;/h2&gt;

&lt;p&gt;The gate is not a separate classifier bolted on the side. It is part of the model's forward pass, and it answers two questions:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Is the model unsure?&lt;/strong&gt; Measured from the entropy and top-probability of the direct pass.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Does this question need step-by-step work?&lt;/strong&gt; Arithmetic, multi-step policy rules, code tracing — the categories where a single forward pass structurally cannot do the job.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Both have to point the same way. Uncertainty alone is not enough: a model can be unsure about a recognition task and thinking will not help, because there is nothing to compute. Structural need alone is not enough either: if the direct pass is already at 97%, the arithmetic is not going to move it.&lt;/p&gt;

&lt;p&gt;When the gate fires, a small adapter writes a thought capped at 96 tokens. Then the model re-reads the state and question with that thought in context and produces a second distribution.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;result&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;decide&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;state&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Applicant: monthly income 4,200 AZN, existing loan payments 1,800 AZN, &lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
          &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;requested new payment 600 AZN.&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;questions&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="nf"&gt;noul&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;dti&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Does this applicant pass the 50% debt-to-income policy?&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)],&lt;/span&gt;
    &lt;span class="n"&gt;think&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;return_thoughts&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="nf"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;result&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;answers&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;].&lt;/span&gt;&lt;span class="n"&gt;probabilities&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="nf"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;result&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;thoughts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;])&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;fail: 75.5%   pass: 24.5%

thought: existing 1800 + new 600 = 2400. 2400 / 4200 = 57.1% &amp;gt; 50% policy limit. fail.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The thought is short on purpose. 96 tokens is enough for arithmetic and a two-step rule check. It is not enough to ramble, and that matters — long reasoning traces are where models go to lose the plot.&lt;/p&gt;




&lt;h2&gt;
  
  
  Why a bad thought cannot overturn a confident answer
&lt;/h2&gt;

&lt;p&gt;This is the part I would want to read if I were evaluating someone else's model.&lt;/p&gt;

&lt;p&gt;The final answer is a weighted mixture of the direct distribution and the post-thought distribution. The weight is not fixed. It is conditioned on how confident the direct pass was:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;If the direct pass is confident, the thought's contribution is bounded. A thought that says "actually it's the other one" cannot drag a 96% answer below the decision threshold.&lt;/li&gt;
&lt;li&gt;If the direct pass is uncertain, the thought carries real weight, because that is exactly the case the gate fired for.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;In practice this means the failure mode is asymmetric in the safe direction. The worst case is that thinking fails to help. The bad case — thinking actively hurts — is suppressed by construction.&lt;/p&gt;

&lt;p&gt;We did not take this on faith. Candidate thoughts were tested on the model itself during training, and kept only if they raised the probability of the correct answer. A thought that did not move the answer was discarded, because we do not pay latency for decoration.&lt;/p&gt;




&lt;h2&gt;
  
  
  Calibration is the actual hard part
&lt;/h2&gt;

&lt;p&gt;Anyone can make a model output numbers between 0 and 1. Making those numbers mean something is the work.&lt;/p&gt;

&lt;p&gt;Three things break naive probability output:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Multi-token labels.&lt;/strong&gt; If your option is "billing_and_refunds", the model tokenizes it into several pieces, and "the probability of the label" becomes ambiguous. THX-02 reads each option with a marker token, so every option is scored through the same path regardless of how many tokens its name contains.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Position bias.&lt;/strong&gt; Options presented in different orders get different scores. We train with option-order averaging, and at inference you can enable &lt;code&gt;tta=4&lt;/code&gt; to average over four orderings. It costs latency; it buys stability.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Calibration drift.&lt;/strong&gt; A model trained with cross-entropy on hard labels learns to be confident, not to be right about its confidence. We trained against full probability distributions with a strictly proper scoring rule plus a KL anchor to the base model. A proper scoring rule is the only family of losses where the optimal strategy is to report your true belief — anything else rewards overconfidence or hedging.&lt;/p&gt;

&lt;p&gt;The practical test is not "is the top answer right." It is: of all the requests where the model said 90%, how many were actually right? If that number is not close to 90, the probabilities are decorative and your thresholds are meaningless.&lt;/p&gt;




&lt;h2&gt;
  
  
  What the gate actually buys, per category
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Category&lt;/th&gt;
&lt;th&gt;Gain from thinking&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Math word problems&lt;/td&gt;
&lt;td&gt;~25 points&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Code tracing&lt;/td&gt;
&lt;td&gt;4–6 points&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Natural language inference&lt;/td&gt;
&lt;td&gt;4–6 points&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Graduate-level science&lt;/td&gt;
&lt;td&gt;4–6 points&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Recognition tasks&lt;/td&gt;
&lt;td&gt;~0&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;That last row is the point. On pure recognition the gate stays off, and you pay nothing. The 25-point category is where a single forward pass is structurally incapable — you cannot compute 2400/4200 without doing the division somewhere.&lt;/p&gt;

&lt;p&gt;If you are building a router, most of your traffic looks like the bottom row. That is why the median stays at 80 ms instead of climbing.&lt;/p&gt;




&lt;h2&gt;
  
  
  Evaluation, honestly
&lt;/h2&gt;

&lt;p&gt;Decision Index 0.3: 65.06 with thinking, 63.83 direct. For context, the best 27B decision model we could find scores 62.25, Cloudflare clef 61.71, Jev 57.96.&lt;/p&gt;

&lt;p&gt;The caveats, because a benchmark number without them is marketing:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;We ran the official kit on a stratified sample of 6,018 requests, not the full suite.&lt;/li&gt;
&lt;li&gt;Competitor numbers are their published public indexes, not our re-runs.&lt;/li&gt;
&lt;li&gt;We decontaminated against test sets: exact and n-gram matching for text, exact and perceptual hashing for all 10,742 evaluation images.&lt;/li&gt;
&lt;li&gt;No benchmark test item was used in training.&lt;/li&gt;
&lt;li&gt;We will submit for independent measurement and publish the result.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The honest summary: the mechanism works, the numbers are ours, and the sample is a sample.&lt;/p&gt;




&lt;h2&gt;
  
  
  Where it fails
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Very large bare-label sets with no descriptions.&lt;/strong&gt; Hand it 200 category names and nothing else and it degrades. Add a one-line description per option and it recovers — the model needs something to score against.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pure recognition.&lt;/strong&gt; Thinking does not help, correctly, but there is also no reasoning gain available.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Long explanations.&lt;/strong&gt; The thought trace is ≤96 tokens by design. If you need a paragraph of justification, this is not the tool.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;thx02
&lt;span class="nb"&gt;export &lt;/span&gt;&lt;span class="nv"&gt;HALX_API_KEY&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"your-key"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The API is free. The weights are Apache-2.0 at &lt;a href="https://huggingface.co/HAL-X-AI/THX-02" rel="noopener noreferrer"&gt;HAL-X-AI/THX-02&lt;/a&gt;, and &lt;code&gt;pip install "thx02[local]"&lt;/code&gt; runs the same model on a single 24 GB GPU if you would rather keep everything in your own perimeter.&lt;/p&gt;

&lt;p&gt;If you build something with the gate, I would like to hear where it fires when it should not. That is the failure mode I am least sure we have fully closed.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;HAL-X, Baku&lt;/strong&gt;&lt;/p&gt;




</description>
      <category>ai</category>
      <category>opensource</category>
      <category>agents</category>
      <category>machinelearning</category>
    </item>
    <item>
      <title>THX-02: a 4.5B open decision model that beats 27B models — and the API is free</title>
      <dc:creator>FX-LgLL</dc:creator>
      <pubDate>Thu, 05 Nov 2026 19:09:13 +0000</pubDate>
      <link>https://dev.to/montydrief/thx-02-a-45b-open-decision-model-that-beats-27b-models-and-the-api-is-free-233n</link>
      <guid>https://dev.to/montydrief/thx-02-a-45b-open-decision-model-that-beats-27b-models-and-the-api-is-free-233n</guid>
      <description>&lt;p&gt;Most of what AI does inside a real product is not writing essays. It is deciding.&lt;/p&gt;

&lt;p&gt;Which team owns this ticket. Is this message phishing. Does this applicant pass the policy. What is the next action. Which of these 40 categories does this document belong to.&lt;/p&gt;

&lt;p&gt;Today teams pick between two bad options. A large chat model: slow, expensive per call, and it will happily invent a label that was never on your list. Or a narrow classifier: fast and cheap, but you retrain it for every new task, and it gives you a score you cannot explain to anyone.&lt;/p&gt;

&lt;p&gt;There is a third path. A model built for decisions only. You send it a state — text, JSON, or up to four images — plus typed questions, and it returns a probability for every option you supplied. It does not generate prose, so it cannot produce an answer outside your list.&lt;/p&gt;

&lt;p&gt;That model is THX-02. 4.5B parameters. Apache-2.0 weights, runtime and server. &lt;strong&gt;And we are giving the API away for free.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;We built it because we needed it ourselves, and we are releasing it because a decision layer that everyone has to pay per token for is a decision layer nobody builds.&lt;/p&gt;




&lt;h2&gt;
  
  
  The free API
&lt;/h2&gt;

&lt;p&gt;This is the part we want to lead with, because it changes who can use this.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Free API access.&lt;/strong&gt; Get a key, call it, ship it. No per-token invoice waiting for you at the end of the month.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Same model as the weights.&lt;/strong&gt; The hosted endpoint runs exactly what you can download. No "API version" that is quietly better than the open one.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No lock-in by design.&lt;/strong&gt; If the free tier ever stops fitting you, &lt;code&gt;pip install "thx02[local]"&lt;/code&gt; puts the same model on one 24 GB GPU, and &lt;code&gt;thx02 serve&lt;/code&gt; gives you the same HTTP interface on your own metal.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Rate limits are published in the API docs&lt;/strong&gt;, so you can size your integration instead of discovering the ceiling in production.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Free is not a marketing word here. The weights are Apache-2.0, so the floor of what you can do without paying anyone is: run it yourself, forever.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;thx02
&lt;span class="nb"&gt;export &lt;/span&gt;&lt;span class="nv"&gt;HALX_API_KEY&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"your-key"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Get a key at the API docs link at the bottom of this post.&lt;/p&gt;




&lt;h2&gt;
  
  
  30 seconds to your first decision
&lt;/h2&gt;

&lt;p&gt;Here is a real request: a customer writes in Azerbaijani that they were charged twice for the same card payment.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;thx02 decide &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--state&lt;/span&gt; &lt;span class="s2"&gt;"Müştəri: kartımdan eyni ödəniş üçün iki dəfə pul çıxıb, 240 AZN. Sifariş #88213."&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--choice&lt;/span&gt; &lt;span class="s2"&gt;"billing,support,shipping,technical"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--yesno&lt;/span&gt; &lt;span class="s2"&gt;"Should this be refunded automatically?"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;billing      ████████████████████████████████  96.0%
support      ██                                   2.4%
shipping     ▏                                    0.9%
technical    ▏                                    0.7%

refund: yes  ███████████████████████████████   95.8%
refund: no   █                                    4.2%

126 ms · think=auto · 1 thought generated
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Two things to notice. The answer is a distribution, not a string — so you can set your own threshold and let the model abstain instead of guessing. And the model decided on its own that this one needed no step-by-step reasoning: it answered directly in 126 ms.&lt;/p&gt;

&lt;p&gt;The same call in Python, with three question types:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;thx02&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;Client&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;choice&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;noul&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;score&lt;/span&gt;

&lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;Client&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;

&lt;span class="n"&gt;result&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;decide&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;state&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Müştəri: kartımdan eyni ödəniş üçün iki dəfə pul çıxıb, 240 AZN. Sifariş #88213.&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;questions&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;
        &lt;span class="nf"&gt;choice&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;team&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;billing&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;support&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;shipping&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;technical&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]),&lt;/span&gt;
        &lt;span class="nf"&gt;noul&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;refund&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Should this be refunded automatically?&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt;
        &lt;span class="nf"&gt;score&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;urgency&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;How urgent is this?&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;5&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt;
    &lt;span class="p"&gt;],&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;q&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;result&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;answers&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="nf"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;q&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nb"&gt;id&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;q&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;probabilities&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;choice&lt;/code&gt; takes up to 255 options. &lt;code&gt;noul&lt;/code&gt; is a yes/no with calibrated probabilities on both sides. &lt;code&gt;score&lt;/code&gt; is an ordered scale, returned as a full distribution over the levels rather than a single number — which means you can see whether the model is confident between 3 and 4, or genuinely torn between 1 and 5.&lt;/p&gt;

&lt;p&gt;Every answer is a full distribution. That is the whole point. A routing system that auto-acts above 0.9 and escalates below 0.6 is a system you can put in front of a customer. A system that returns &lt;code&gt;"billing"&lt;/code&gt; and nothing else is a system you have to babysit.&lt;/p&gt;




&lt;h2&gt;
  
  
  Thinking, but only when it changes the answer
&lt;/h2&gt;

&lt;p&gt;The interesting engineering problem in a decision model is not "make it smarter." It is "make it smarter without making it slower and without letting it talk itself out of a correct answer."&lt;/p&gt;

&lt;p&gt;THX-02 has a built-in gate:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The model answers directly first.&lt;/li&gt;
&lt;li&gt;The gate asks two questions: is the model unsure, and does this question need step-by-step work — arithmetic, multi-step rules, code tracing.&lt;/li&gt;
&lt;li&gt;Only if both point that way does a small adapter write a short thought, capped at 96 tokens.&lt;/li&gt;
&lt;li&gt;The model re-reads the question with that thought and mixes the two distributions, weighted so a bad thought cannot overturn a confident direct answer.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;That last step is the important one. Thinking is an additive signal, not a replacement. If the direct pass is confident and the thought disagrees, the thought loses.&lt;/p&gt;

&lt;p&gt;Median latency stays around 80 ms with the gate active. On math word problems the thought adds about 25 points. On natural language inference, code tracing and graduate-level science it adds 4 to 6 — real, but not worth paying 10x latency for on every request, which is exactly why the gate exists.&lt;/p&gt;

&lt;p&gt;Here is a case where it fires. A policy check that requires arithmetic:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;result&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;decide&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;state&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Applicant: monthly income 4,200 AZN, existing loan payments 1,800 AZN, &lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
          &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;requested new payment 600 AZN.&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;questions&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="nf"&gt;noul&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;dti&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Does this applicant pass the 50% debt-to-income policy?&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)],&lt;/span&gt;
    &lt;span class="n"&gt;think&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;return_thoughts&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="nf"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;result&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;answers&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;].&lt;/span&gt;&lt;span class="n"&gt;probabilities&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="nf"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;result&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;thoughts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;])&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;fail: 75.5%   pass: 24.5%

thought: existing 1800 + new 600 = 2400. 2400 / 4200 = 57.1% &amp;gt; 50% policy limit. fail.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The model did not guess. It computed, showed the arithmetic, and the probability moved because the arithmetic moved it.&lt;/p&gt;




&lt;h2&gt;
  
  
  Results
&lt;/h2&gt;

&lt;p&gt;Decision Index 0.3, the open benchmark suite for decision models:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;Size&lt;/th&gt;
&lt;th&gt;Decision Index 0.3&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;THX-02 (thinking)&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;4.5B&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;65.06&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;THX-02 (direct)&lt;/td&gt;
&lt;td&gt;4.5B&lt;/td&gt;
&lt;td&gt;63.83&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Best 27B decision model&lt;/td&gt;
&lt;td&gt;27B&lt;/td&gt;
&lt;td&gt;62.25&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Cloudflare clef&lt;/td&gt;
&lt;td&gt;—&lt;/td&gt;
&lt;td&gt;61.71&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Jev&lt;/td&gt;
&lt;td&gt;—&lt;/td&gt;
&lt;td&gt;57.96&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;THX-01&lt;/td&gt;
&lt;td&gt;0.3B&lt;/td&gt;
&lt;td&gt;11.30&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Read that table again. A 4.5B model scores above the best 27B decision model we could find, and above Jev's published index by more than seven points. On the public Decision Vision Index sets it reaches 72.19.&lt;/p&gt;

&lt;p&gt;We also want to be precise about what those numbers are, because "we beat X" deserves the methodology attached:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;We ran the official evaluation kit ourselves, on a stratified sample of 6,018 requests rather than the full suite.&lt;/li&gt;
&lt;li&gt;The competitor numbers are their published public indexes, not our re-runs.&lt;/li&gt;
&lt;li&gt;We decontaminated against the test sets: exact and n-gram matching for text, exact and perceptual hashing for all 10,742 evaluation images.&lt;/li&gt;
&lt;li&gt;No benchmark test item was used in training.&lt;/li&gt;
&lt;li&gt;We will submit to the maintainers for an independent measurement, and we will publish whatever comes back.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If the independent number lands lower, we will say so in the same place we said 65.06.&lt;/p&gt;




&lt;h2&gt;
  
  
  How we got there
&lt;/h2&gt;

&lt;p&gt;Five post-training stages on top of an open base model. Each one was kept only if it improved a held-out split — no stage survived on vibes.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Distribution distillation.&lt;/strong&gt; 361,000 requests, trained against full probability distributions with a strictly proper scoring rule plus a KL anchor to the base model. Proper scoring rules are what make the output calibrated instead of merely confident.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Disagreement mining.&lt;/strong&gt; A pass over cases where the teacher and the student disagreed, keeping the ones where the teacher was right and the student was confidently wrong.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Inference-structure training.&lt;/strong&gt; Option-order averaging, joint reading of state and question, and marker-token scoring per option — the things that remove position bias and multi-token label tokenization problems you would otherwise have to solve yourself.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Self-verified thoughts.&lt;/strong&gt; Candidate thoughts were tested on the model itself and kept only if they raised the probability of the correct answer. A thought that does not change the answer is a thought we do not pay latency for.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Targeted skills.&lt;/strong&gt; About 700,000 candidate examples, with answers verified by executing code, running solvers or querying engines rather than asking another model. 161,000 survived — the ones the model still missed or was unsure about.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;That last point matters more than it sounds. A model trained on another model's answers inherits that model's mistakes, confidently. A model trained on executed code and solver output inherits the truth.&lt;/p&gt;




&lt;h2&gt;
  
  
  Run it anywhere
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Hosted, free:&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;thx02
&lt;span class="nb"&gt;export &lt;/span&gt;&lt;span class="nv"&gt;HALX_API_KEY&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"your-key"&lt;/span&gt;
thx02 decide &lt;span class="nt"&gt;--state&lt;/span&gt; &lt;span class="s2"&gt;"..."&lt;/span&gt; &lt;span class="nt"&gt;--choice&lt;/span&gt; &lt;span class="s2"&gt;"a,b,c"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;On your own GPU&lt;/strong&gt; — one 24 GB card is enough:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="s2"&gt;"thx02[local]"&lt;/span&gt;
python &lt;span class="nt"&gt;-c&lt;/span&gt; &lt;span class="s2"&gt;"from thx02 import THX02; THX02.load('HAL-X-AI/THX-02')"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;As a server&lt;/strong&gt; with the same HTTP interface as the hosted API:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="s2"&gt;"thx02[server]"&lt;/span&gt;
thx02 serve
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;On a phone.&lt;/strong&gt; The client is zero-dependency and runs in Termux:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pkg &lt;span class="nb"&gt;install &lt;/span&gt;python
pip &lt;span class="nb"&gt;install &lt;/span&gt;thx02
&lt;span class="nb"&gt;export &lt;/span&gt;&lt;span class="nv"&gt;HALX_API_KEY&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"your-key"&lt;/span&gt;
thx02 decide &lt;span class="nt"&gt;--state&lt;/span&gt; &lt;span class="s2"&gt;"..."&lt;/span&gt; &lt;span class="nt"&gt;--yesno&lt;/span&gt; &lt;span class="s2"&gt;"Is this message spam?"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Fine-tuned to your domain.&lt;/strong&gt; The weights load through ms-swift like any other model of this family, and the think adapter is a separate artifact you can keep, drop or retrain.&lt;/p&gt;

&lt;p&gt;CPU inference is possible but not the point: about 0.6 s per short request on 8 x86 cores in BF16. On a GPU you are looking at 50 to 150 ms.&lt;/p&gt;




&lt;h2&gt;
  
  
  What it is not good at
&lt;/h2&gt;

&lt;p&gt;We would rather tell you now than have you find out in production.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Very large bare-label sets with no descriptions.&lt;/strong&gt; If you hand it 200 category names and nothing else, it degrades. Give each option a short description and it recovers.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pure recognition tasks.&lt;/strong&gt; Thinking does not help much on "what is in this image" style questions — the gate correctly stays off, but there is no reasoning gain to be had.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Anything that needs a long generation.&lt;/strong&gt; It is a decision model. It returns distributions, not paragraphs. If you need an explanation, the thought trace is short by design (≤96 tokens).&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Why open
&lt;/h2&gt;

&lt;p&gt;A decision layer sits in the middle of your product. It sees customer messages, applicant data, medical codes, transaction descriptions. For anyone in a regulated sector — banking, insurance, health, public sector — "send it to a remote black box and hope" is not an architecture, it is a risk.&lt;/p&gt;

&lt;p&gt;So: Apache-2.0 weights, the think adapter, the runtime and the server. On-premise, inspectable, adaptable. Multilingual from launch, with English, Russian, Chinese, Turkish and Spanish alongside a deliberate focus on Azerbaijani, Kazakh, Uzbek and Kyrgyz — languages that most decision models treat as an afterthought, if they treat them at all.&lt;/p&gt;

&lt;p&gt;And free to call, because the fastest way to find out whether a decision model belongs in your product is to put one in your product this afternoon.&lt;/p&gt;




&lt;h2&gt;
  
  
  Links
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Weights: &lt;a href="https://huggingface.co/HAL-X-AI/THX-02" rel="noopener noreferrer"&gt;HAL-X-AI/THX-02&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Client and CLI: &lt;code&gt;pip install thx02&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;Free API key and docs: see the API section of the model card&lt;/li&gt;
&lt;li&gt;Local runtime: &lt;code&gt;pip install "thx02[local]"&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;Server: &lt;code&gt;pip install "thx02[server]"&lt;/code&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;We can't wait to see what you decide with it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;HAL-X, Baku&lt;/strong&gt;&lt;/p&gt;




</description>
      <category>ai</category>
      <category>opensource</category>
      <category>api</category>
      <category>machinelearning</category>
    </item>
    <item>
      <title>THX-02: an open 4.5B decision model that thinks only when it changes the answer</title>
      <dc:creator>FX-LgLL</dc:creator>
      <pubDate>Thu, 05 Nov 2026 18:13:44 +0000</pubDate>
      <link>https://dev.to/montydrief/thx-02-an-open-45b-decision-model-that-thinks-only-when-it-changes-the-answer-bom</link>
      <guid>https://dev.to/montydrief/thx-02-an-open-45b-decision-model-that-thinks-only-when-it-changes-the-answer-bom</guid>
      <description>&lt;p&gt;Most of what AI does inside a real product is not writing essays. It is &lt;strong&gt;deciding&lt;/strong&gt;.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Which team should handle this ticket?&lt;/li&gt;
&lt;li&gt;Is this email phishing?&lt;/li&gt;
&lt;li&gt;Does this loan application pass policy?&lt;/li&gt;
&lt;li&gt;Which button should the agent click next?&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Today you usually solve that in one of two ways. You prompt a large chat model and parse its text: slow, expensive and&lt;br&gt;
sometimes it invents a label you never offered. Or you train a small classifier for one fixed label set, which is fast&lt;br&gt;
but has to be retrained for every new task.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;THX-02 is a third option.&lt;/strong&gt; It is a decision model: you send a state and typed questions, and it returns a calibrated&lt;br&gt;
probability for every option you gave it. When a question needs a calculation or a chain of logic, it writes a short&lt;br&gt;
thought first. When it doesn't, it doesn't.&lt;/p&gt;

&lt;p&gt;Today we are releasing THX-02 as open source under Apache-2.0.&lt;/p&gt;
&lt;h2&gt;
  
  
  30 seconds to your first decision
&lt;/h2&gt;


&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;thx02
&lt;span class="nb"&gt;export &lt;/span&gt;&lt;span class="nv"&gt;HALX_API_KEY&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;sk-...
thx02 &lt;span class="s2"&gt;"Salam, kartımdan iki dəfə pul çıxılıb, geri qaytarın"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--choice&lt;/span&gt; &lt;span class="s2"&gt;"team=billing:charges and refunds,support:technical problems,sales:buying a product"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--yesno&lt;/span&gt; &lt;span class="s2"&gt;"Does the customer ask for a refund?"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;team
  › billing                  ███████████████████████   96.0%
    support                  █                          3.5%
    sales                                               0.4%

Does the customer ask for a refund
  › yes                      ███████████████████████   95.8%
    no                       █                          4.2%

THX-02 · 126 ms
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;The same thing from Python:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;thx02&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;Client&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;choice&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;noul&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;score&lt;/span&gt;

&lt;span class="n"&gt;thx&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;Client&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;
&lt;span class="n"&gt;out&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;thx&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;decide&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Salam, kartımdan iki dəfə pul çıxılıb, geri qaytarın zəhmət olmasa.&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;team&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;   &lt;span class="nf"&gt;choice&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Which team should handle this?&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;billing&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Charges and refunds&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                         &lt;span class="n"&gt;support&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Technical problems&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;sales&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Buying a product&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;refund&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nf"&gt;noul&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Does the customer ask for a refund?&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;anger&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;  &lt;span class="nf"&gt;score&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;How frustrated is the customer?&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;calm&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;annoyed&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;angry&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]),&lt;/span&gt;
    &lt;span class="p"&gt;},&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;There are three question types: &lt;code&gt;choice&lt;/code&gt; (one of up to 255 options), &lt;code&gt;noul&lt;/code&gt; (yes / no) and &lt;code&gt;score&lt;/code&gt; (an ordered&lt;br&gt;
scale). Every answer is a full probability distribution, so you can set thresholds: act automatically above 0.9, send&lt;br&gt;
the rest to a person.&lt;/p&gt;

&lt;h2&gt;
  
  
  Thinking, but only when it helps
&lt;/h2&gt;

&lt;p&gt;Most decisions are recognition: the answer is visible in the text. Some need work first. For those, THX-02:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;answers every question directly;&lt;/li&gt;
&lt;li&gt;checks, with a built-in gate, whether it is unsure &lt;em&gt;and&lt;/em&gt; whether the question needs step-by-step work;&lt;/li&gt;
&lt;li&gt;only then writes a short thought (up to 96 tokens) with the facts or the calculation;&lt;/li&gt;
&lt;li&gt;re-reads the question with the thought and mixes both reads. A wrong thought cannot overturn a confident answer.
&lt;/li&gt;
&lt;/ol&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;out&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;thx&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;decide&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Applicant income $4,200/month; rent $1,500; car loan $450; cards $250; new loan payment $600. &lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Policy: total debt payments (including rent) must not exceed 50% of income.&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;policy&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nf"&gt;choice&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Does the application pass the bank policy?&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;passes&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Within the limit&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;fails&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Over the limit&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)},&lt;/span&gt;
    &lt;span class="n"&gt;think&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;return_thoughts&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;verdict
  › fail                     ██████████████████        75.5%
    pass                     ██████                    24.5%
  thought: Total monthly debt payments = $1,500 (rent) + $450 (car) + $250 (cards) + $600 (new loan) = $2,800.
           50% of income = $4,200 × 0.50 = $2,100. Since $2,800 &amp;gt; $2,100, the application exceeds the 50% limit.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Because thinking is gated, the median request with &lt;code&gt;think="auto"&lt;/code&gt; still takes about 80 ms. On math word problems&lt;br&gt;
thinking adds 25 points; on code tracing, natural-language inference and graduate-level science it adds 4 to 6.&lt;/p&gt;

&lt;h2&gt;
  
  
  Results
&lt;/h2&gt;

&lt;p&gt;Decision Index 0.3 covers 37 decision tasks in five areas: knowledge and reasoning, language understanding, retrieval and&lt;br&gt;
classification, tools and automation, and arts and human taste.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;Size&lt;/th&gt;
&lt;th&gt;Decision Index 0.3&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;THX-02, thinking&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;4.5 B&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;65.06&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;THX-02, direct&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;4.5 B&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;63.83&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Best 27B decision model on the board&lt;/td&gt;
&lt;td&gt;27 B&lt;/td&gt;
&lt;td&gt;62.25&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Cloudflare clef&lt;/td&gt;
&lt;td&gt;&lt;/td&gt;
&lt;td&gt;61.71&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Jev&lt;/td&gt;
&lt;td&gt;&lt;/td&gt;
&lt;td&gt;57.96&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;THX-01 (our previous model)&lt;/td&gt;
&lt;td&gt;0.3 B&lt;/td&gt;
&lt;td&gt;11.30&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;On the public sets of the Decision Vision Index, THX-02 scores &lt;strong&gt;72.19&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;How we measured:&lt;/em&gt; we ran THX-02 with the official Decision Index 0.3 kit on a stratified sample of 6,018 requests&lt;br&gt;
from the public suite. The other scores are the published public indexes. No benchmark test item was used in training.&lt;br&gt;
Every training text was checked against the test sets with exact and n-gram matching, and every image against all&lt;br&gt;
10,742 evaluation images with exact and perceptual hashes.&lt;/p&gt;

&lt;h2&gt;
  
  
  How we got there
&lt;/h2&gt;

&lt;p&gt;THX-02 was post-trained in five stages, and a stage stayed only if it raised the score on a held-out split:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Distillation:&lt;/strong&gt; a larger decision model labelled 361,000 requests with full probability distributions. The
objective is a strictly proper scoring rule, plus a KL anchor so the model keeps what it already knew.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Disagreements:&lt;/strong&gt; we kept the requests where student and teacher disagreed &lt;em&gt;and&lt;/em&gt; the teacher was right.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Inference structure:&lt;/strong&gt; choice options are read in several orders and averaged; related questions are read
together.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Learning to think:&lt;/strong&gt; candidate thoughts were tested on THX-02 itself and kept only if they made the right answer
more likely.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Targeted skills:&lt;/strong&gt; about 700,000 new candidates, with answers verified by running code wherever possible
(sandboxed Python, solvers, a chess engine, simulations). We kept the 161,000 the model still got wrong or was
unsure about.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Run it anywhere
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;API:&lt;/strong&gt; &lt;code&gt;pip install thx02&lt;/code&gt;, then &lt;code&gt;Client()&lt;/code&gt;. Reference: &lt;a href="https://api.hal-x.ai/docs/thx-02" rel="noopener noreferrer"&gt;https://api.hal-x.ai/docs/thx-02&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Your GPU:&lt;/strong&gt; &lt;code&gt;pip install "thx02[local]"&lt;/code&gt;, then &lt;code&gt;THX02.load("HAL-X-AI/THX-02")&lt;/code&gt;. One RTX 4090 is enough.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Your own server:&lt;/strong&gt; &lt;code&gt;pip install "thx02[server]"&lt;/code&gt;, then &lt;code&gt;thx02 serve&lt;/code&gt;. It speaks the same API.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Your phone:&lt;/strong&gt; the client has no dependencies, so in Termux it's &lt;code&gt;pkg install python &amp;amp;&amp;amp; pip install thx02&lt;/code&gt;. Keep
an interactive session alive in tmux.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Fine-tuning:&lt;/strong&gt; supported in ms-swift with &lt;code&gt;--model_type thx02&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Why open
&lt;/h2&gt;

&lt;p&gt;Teams that put AI into banks, hospitals, factories and public services need to run it on their own hardware, inspect it&lt;br&gt;
and adapt it. A decision that blocks a payment or routes a patient should not depend on a black box behind someone&lt;br&gt;
else's API. THX-02 is Apache-2.0, multilingual from day one, and built with particular care for languages that are&lt;br&gt;
usually left behind: Azerbaijani, Kazakh, Uzbek and Kyrgyz, alongside English, Russian, Chinese, Turkish and Spanish.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Weights: &lt;a href="https://huggingface.co/HAL-X-AI/THX-02" rel="noopener noreferrer"&gt;https://huggingface.co/HAL-X-AI/THX-02&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;API docs: &lt;a href="https://api.hal-x.ai/docs/thx-02" rel="noopener noreferrer"&gt;https://api.hal-x.ai/docs/thx-02&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Install: &lt;code&gt;pip install thx02&lt;/code&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;We can't wait to see what you decide with it.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;HAL-X, Baku&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>opensource</category>
      <category>machinelearning</category>
      <category>llm</category>
    </item>
    <item>
      <title>THX-01 Just Processed 70 Million Tokens. Our Open-Source Decision Model Is Just Getting Started.</title>
      <dc:creator>FX-LgLL</dc:creator>
      <pubDate>Mon, 12 Oct 2026 09:36:54 +0000</pubDate>
      <link>https://dev.to/montydrief/thx-01-just-processed-70-million-tokens-our-open-source-decision-model-is-just-getting-started-2o8l</link>
      <guid>https://dev.to/montydrief/thx-01-just-processed-70-million-tokens-our-open-source-decision-model-is-just-getting-started-2o8l</guid>
      <description>&lt;p&gt;&lt;strong&gt;70 million tokens processed. 322 million parameters. One forward pass. Free API. Fully open-source.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;A few days ago, we open-sourced &lt;strong&gt;THX-01&lt;/strong&gt;, our 322M-parameter non-autoregressive decision model designed to make AI decisions faster, more efficient, and accessible to everyone.&lt;/p&gt;

&lt;p&gt;Today, we've reached another milestone: &lt;strong&gt;over 70 million tokens processed through our THX-01 API since launch.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;And we're just getting started.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why THX-01?
&lt;/h2&gt;

&lt;p&gt;Most AI applications don't actually need to generate long responses.&lt;/p&gt;

&lt;p&gt;Sometimes, all you need is a reliable decision:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Which department should handle this ticket?&lt;/li&gt;
&lt;li&gt;Is this message spam?&lt;/li&gt;
&lt;li&gt;What number appears in this document?&lt;/li&gt;
&lt;li&gt;Which passage supports the answer?&lt;/li&gt;
&lt;li&gt;How confident is the model in its decision?&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Instead of generating tokens one by one, THX-01 processes a document or message alongside typed questions and returns structured answers with calibrated probabilities in a single forward pass.&lt;/p&gt;

&lt;p&gt;On our GPU benchmark, inference takes approximately &lt;strong&gt;10 milliseconds&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  THX-01 vs TypeSafe Jev
&lt;/h2&gt;

&lt;p&gt;We evaluated THX-01 against TypeSafe Jev 1.13 on our internal benchmark of 2,843 multilingual support tickets.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Metric&lt;/th&gt;
&lt;th&gt;THX-01&lt;/th&gt;
&lt;th&gt;Jev 1.13&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Accuracy&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;98.4%&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;97.4%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Calibration Error (ECE)&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;0.003&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;0.007&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Measured Latency&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;~10 ms&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;331 ms&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Open-Source Weights&lt;/td&gt;
&lt;td&gt;Yes&lt;/td&gt;
&lt;td&gt;No&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;THX-01 outperformed Jev on this classification benchmark, with approximately 33x lower measured latency.&lt;/p&gt;

&lt;p&gt;The latency figures come from different serving environments, so they should not be interpreted as a hardware-controlled comparison.&lt;/p&gt;

&lt;p&gt;Our results show that small, specialized decision models can deliver competitive performance without requiring billions of additional parameters.&lt;/p&gt;

&lt;h2&gt;
  
  
  Built for Efficient Decision-Making
&lt;/h2&gt;

&lt;p&gt;THX-01 is trained using &lt;strong&gt;Reinforcement Learning for Calibrated Decisions (RLCD)&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Rather than optimizing only for correct predictions, RLCD encourages probability estimates that reflect the model's actual uncertainty.&lt;/p&gt;

&lt;p&gt;This is particularly useful for AI agents and production applications that need to decide whether to accept a prediction, request additional information, or escalate to another model.&lt;/p&gt;

&lt;p&gt;THX-01 supports:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Choice:&lt;/strong&gt; Classification among dynamically provided options&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Yes/No:&lt;/strong&gt; Binary decisions with confidence scores&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Score:&lt;/strong&gt; Ordered evaluations&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Number:&lt;/strong&gt; Extracting numerical values from documents&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Excerpt:&lt;/strong&gt; Extracting exact text passages&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Citations:&lt;/strong&gt; Identifying supporting evidence&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The model supports 18 post-training languages, including Azerbaijani, English, Russian, and Turkish.&lt;/p&gt;

&lt;p&gt;And at just 322M parameters, it can even run locally on smartphones with a compatible inference runtime.&lt;/p&gt;

&lt;h2&gt;
  
  
  The THX-01 API Is Free
&lt;/h2&gt;

&lt;p&gt;We believe developers should be able to experiment with new AI architectures without worrying about infrastructure costs.&lt;/p&gt;

&lt;p&gt;That's why we've opened the &lt;strong&gt;THX-01 API with a free tier of 200 decisions per minute&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;You can test it with a simple request:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl https://api.hal-x.ai/v1/systemone &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s2"&gt;"Content-Type: application/json"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-d&lt;/span&gt; &lt;span class="s1"&gt;'{
    "state": "My card was charged twice for one order",
    "questions": {
      "team": {
        "type": "choice",
        "question": "Which team should handle this?",
        "criteria": {
          "billing": "Billing issues",
          "technical": "Technical issues",
          "sales": "Sales inquiries"
        }
      }
    }
  }'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Check the API documentation for the current authentication requirements and usage limits.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;API Documentation:&lt;/strong&gt; &lt;a href="https://api.hal-x.ai/docs/thx-01" rel="noopener noreferrer"&gt;https://api.hal-x.ai/docs/thx-01&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Everything Is Open-Source
&lt;/h2&gt;

&lt;p&gt;We released THX-01 under the &lt;strong&gt;Apache 2.0 license&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Developers can access the model weights, source code, and training resources on Hugging Face.&lt;/p&gt;

&lt;p&gt;You can also install the Python package directly:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;thx01
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Model &amp;amp; Source Code:&lt;/strong&gt; &lt;a href="https://huggingface.co/doofz/THX-01" rel="noopener noreferrer"&gt;https://huggingface.co/doofz/THX-01&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Interactive Demo:&lt;/strong&gt; &lt;a href="https://huggingface.co/spaces/doofz/THX-01-demo" rel="noopener noreferrer"&gt;https://huggingface.co/spaces/doofz/THX-01-demo&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What's Next?
&lt;/h2&gt;

&lt;p&gt;At HAL-X AI, we're working toward a future where AI systems don't just generate answers, but make reliable, measurable, and well-calibrated decisions.&lt;/p&gt;

&lt;p&gt;We believe the next generation of AI infrastructure will need specialized models that are smaller, faster, and capable of expressing uncertainty.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;70 million tokens processed is an exciting milestone for our team, but this is only the beginning.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;We're continuing to improve THX-01, expand its capabilities, and contribute to the global open-source AI ecosystem.&lt;/p&gt;

&lt;p&gt;Our mission is simple:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;We need more AI models that know how to make decisions, how confident they should be, and when they don't know.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;And we're building them for everyone.&lt;/p&gt;




&lt;p&gt;&lt;strong&gt;Try THX-01 today:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Free API: &lt;a href="https://api.hal-x.ai/docs/thx-01" rel="noopener noreferrer"&gt;https://api.hal-x.ai/docs/thx-01&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Hugging Face: &lt;a href="https://huggingface.co/doofz/THX-01" rel="noopener noreferrer"&gt;https://huggingface.co/doofz/THX-01&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Python: &lt;code&gt;pip install thx01&lt;/code&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>opensource</category>
      <category>machinelearning</category>
      <category>api</category>
    </item>
    <item>
      <title>We Open-Sourced a 322M Decision Model. Now Its API Is Free</title>
      <dc:creator>FX-LgLL</dc:creator>
      <pubDate>Sat, 10 Oct 2026 21:34:17 +0000</pubDate>
      <link>https://dev.to/montydrief/we-open-sourced-a-322m-decision-model-now-its-api-is-free-187e</link>
      <guid>https://dev.to/montydrief/we-open-sourced-a-322m-decision-model-now-its-api-is-free-187e</guid>
      <description>&lt;p&gt;&lt;strong&gt;Two days ago, we released THX-01, an open-source decision model that outperforms Jev on several benchmarks and can even run locally on a smartphone. Today, anyone can use it through our free API.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Hi everyone! I'm Farid, Co-Founder of HAL-X AI.&lt;/p&gt;

&lt;p&gt;A couple of days ago, our team open-sourced &lt;strong&gt;THX-01&lt;/strong&gt;, a 322M-parameter multilingual decision model designed to make AI decision-making faster, cheaper, and more accessible.&lt;/p&gt;

&lt;p&gt;Yesterday, we opened our hosted API to everyone. &lt;strong&gt;No API key, no sign-up, and no payment required.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;And the response has been incredible!&lt;/p&gt;

&lt;p&gt;Since launch, we've already processed:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;25M+ tokens&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;211,000+ decisions&lt;/strong&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;We're excited to see developers experimenting with what a small, specialized decision model can do.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Is THX-01?
&lt;/h2&gt;

&lt;p&gt;Most applications don't actually need an LLM to generate paragraphs of text.&lt;/p&gt;

&lt;p&gt;Sometimes, you just need to answer a simple question:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Is this email spam?&lt;/li&gt;
&lt;li&gt;Which department should handle this support ticket?&lt;/li&gt;
&lt;li&gt;What is the invoice total?&lt;/li&gt;
&lt;li&gt;Does this document mention a particular condition?&lt;/li&gt;
&lt;li&gt;Which passage supports this decision?&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Using a large generative model for these tasks can introduce unnecessary latency and cost.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;THX-01 takes a different approach.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;It's a non-autoregressive model that processes a document or message alongside typed questions and returns structured answers with calibrated probabilities in a single forward pass.&lt;/p&gt;

&lt;p&gt;Approximately &lt;strong&gt;10 milliseconds&lt;/strong&gt; on our benchmark hardware.&lt;/p&gt;

&lt;p&gt;No token-by-token generation.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try the Free API
&lt;/h2&gt;

&lt;p&gt;You can make your first request right now.&lt;/p&gt;

&lt;p&gt;No authentication required.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl https://api.hal-x.ai/v1/systemone &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s2"&gt;"Content-Type: application/json"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-d&lt;/span&gt; &lt;span class="s1"&gt;'{
    "state": "My card was charged twice for one order",
    "questions": {
      "team": {
        "type": "choice",
        "question": "Which team handles this?",
        "criteria": {
          "billing": "billing",
          "tech": "technical",
          "sales": "sales"
        }
      }
    }
  }'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;THX-01 evaluates the available choices and returns the predicted decision with probabilities.&lt;/p&gt;

&lt;p&gt;The API supports:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Type&lt;/th&gt;
&lt;th&gt;What it does&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Choice&lt;/td&gt;
&lt;td&gt;Classification with probabilities&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Yes/No&lt;/td&gt;
&lt;td&gt;Binary decisions&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Score&lt;/td&gt;
&lt;td&gt;Ordered scoring&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Number&lt;/td&gt;
&lt;td&gt;Extract exact numbers from documents&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Excerpt&lt;/td&gt;
&lt;td&gt;Extract verbatim passages&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Citations&lt;/td&gt;
&lt;td&gt;Identify supporting text&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;You can submit multiple questions in a single request.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Full API documentation:&lt;/strong&gt; &lt;a href="https://api.hal-x.ai/docs/thx-01" rel="noopener noreferrer"&gt;https://api.hal-x.ai/docs/thx-01&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  How Does It Compare to Jev?
&lt;/h2&gt;

&lt;p&gt;We benchmarked THX-01 against TypeSafe Jev 1.13 using 2,843 multilingual support tickets across four languages.&lt;/p&gt;

&lt;p&gt;Here are the results:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Metric&lt;/th&gt;
&lt;th&gt;THX-01&lt;/th&gt;
&lt;th&gt;Jev 1.13&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Accuracy&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;98.4%&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;97.4%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Calibration error&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;0.003&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;0.007&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Measured latency&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;~10 ms&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;331 ms&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Open weights&lt;/td&gt;
&lt;td&gt;Yes&lt;/td&gt;
&lt;td&gt;No&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;THX-01 achieved higher accuracy and better calibration on this benchmark, with approximately 33x lower measured latency.&lt;/p&gt;

&lt;p&gt;The latency measurements use different serving setups, so they're not a controlled hardware-to-hardware comparison.&lt;/p&gt;

&lt;p&gt;We don't claim THX-01 outperforms Jev everywhere, but these results demonstrate how competitive specialized, lightweight decision models can be.&lt;/p&gt;

&lt;h2&gt;
  
  
  Small Enough to Run on Your Phone
&lt;/h2&gt;

&lt;p&gt;THX-01 contains just &lt;strong&gt;322 million parameters&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;It was designed to support efficient inference without requiring a massive GPU infrastructure.&lt;/p&gt;

&lt;p&gt;With a suitable runtime, it can even run locally on smartphones.&lt;/p&gt;

&lt;p&gt;You can also install the Python package:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;thx01
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The model supports 18 post-training languages, with particular attention to Azerbaijani and multilingual robustness.&lt;/p&gt;

&lt;h2&gt;
  
  
  Everything Is Open-Source
&lt;/h2&gt;

&lt;p&gt;We're releasing THX-01 under &lt;strong&gt;Apache 2.0&lt;/strong&gt;, including its model weights, source code, and training scripts.&lt;/p&gt;

&lt;p&gt;We believe developers should have the freedom to run, modify, deploy, and build on top of their AI models.&lt;/p&gt;

&lt;p&gt;At HAL-X AI, our goal is to contribute to the global open-source AI ecosystem and make efficient AI accessible to more people.&lt;/p&gt;

&lt;p&gt;That's also why we've made the hosted API free to use, subject to rate limits.&lt;/p&gt;

&lt;h2&gt;
  
  
  Get Started
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Free API:&lt;/strong&gt; &lt;a href="https://api.hal-x.ai/docs/thx-01" rel="noopener noreferrer"&gt;https://api.hal-x.ai/docs/thx-01&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Model weights and code:&lt;/strong&gt; &lt;a href="https://huggingface.co/doofz/THX-01" rel="noopener noreferrer"&gt;https://huggingface.co/doofz/THX-01&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Interactive demo:&lt;/strong&gt; &lt;a href="https://huggingface.co/spaces/doofz/THX-01-demo" rel="noopener noreferrer"&gt;https://huggingface.co/spaces/doofz/THX-01-demo&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Python package:&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;thx01
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;We're still improving THX-01, and we'd love to hear feedback from developers building AI agents, automation pipelines, document processing systems, and classification services.&lt;/p&gt;

&lt;p&gt;If you build something using THX-01, we'd love to see it!&lt;/p&gt;

</description>
      <category>ai</category>
      <category>opensource</category>
      <category>api</category>
      <category>machinelearning</category>
    </item>
    <item>
      <title>We made our decision model's API free (and the weights are open)</title>
      <dc:creator>FX-LgLL</dc:creator>
      <pubDate>Fri, 09 Oct 2026 19:43:18 +0000</pubDate>
      <link>https://dev.to/montydrief/we-made-our-decision-models-api-free-and-the-weights-are-open-5c6n</link>
      <guid>https://dev.to/montydrief/we-made-our-decision-models-api-free-and-the-weights-are-open-5c6n</guid>
      <description>&lt;p&gt;Most "AI decisions" in production aren't open-ended generation. They're classification in disguise:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;which team should handle this ticket?&lt;/li&gt;
&lt;li&gt;is this e-mail phishing?&lt;/li&gt;
&lt;li&gt;what's the total on this invoice?&lt;/li&gt;
&lt;li&gt;which sentence in the contract supports that answer?&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Teams usually send each of these to a large LLM: 1–2 seconds and a bill per call, a JSON answer to parse, and a "confidence" that means nothing.&lt;/p&gt;

&lt;p&gt;We built &lt;strong&gt;THX-01&lt;/strong&gt; for exactly these decisions. Today we're making its hosted API free.&lt;/p&gt;

&lt;h3&gt;
  
  
  Try it now (no key, no sign-up)
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl https://api.hal-x.ai/v1/systemone &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s2"&gt;"Content-Type: application/json"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-d&lt;/span&gt; &lt;span class="s1"&gt;'{
    "state": "Hi, I was charged twice for one order and nobody answers. Refund me!",
    "questions": {
      "team":   {"type": "choice", "question": "Which team handles this?",
                 "criteria": {"billing": "billing", "tech": "technical", "sales": "sales"}},
      "refund": {"type": "noul", "question": "Does the customer want money back?"},
      "urgency":{"type": "score", "question": "How urgent?", "criteria": ["low", "medium", "high"]}
    }
  }'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Every answer comes back with a probability for each option, computed in a single forward pass of about 10 ms.&lt;/p&gt;

&lt;p&gt;The free tier allows 200 decisions per minute per IP. One question is one decision.&lt;/p&gt;

&lt;h3&gt;
  
  
  What it can answer
&lt;/h3&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;type&lt;/th&gt;
&lt;th&gt;returns&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;choice&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;one option, with a probability for every option&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;noul&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;P(yes)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;score&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;a level on an ordinal scale&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;number&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;a value stated in the document, or &lt;code&gt;null&lt;/code&gt; if it isn't there&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;excerpt&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;a verbatim span with character offsets&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;"cite": true&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;the passages that support any answer&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h3&gt;
  
  
  How it compares
&lt;/h3&gt;

&lt;p&gt;On our 2,843-ticket benchmark (Azerbaijani, Russian, English, Turkish; clean, corrupted, messy and independently written sets):&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;model&lt;/th&gt;
&lt;th&gt;avg accuracy&lt;/th&gt;
&lt;th&gt;latency&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;THX-01 (322M, open)&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;98.4&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;~10 ms&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;TypeSafe Jev 1.13&lt;/td&gt;
&lt;td&gt;97.4&lt;/td&gt;
&lt;td&gt;331 ms&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Kev-4B&lt;/td&gt;
&lt;td&gt;92.8&lt;/td&gt;
&lt;td&gt;830 ms&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The model's confidence is calibrated, so a single threshold lets it handle about 75% of tickets automatically with zero errors on all four test sets.&lt;/p&gt;

&lt;h3&gt;
  
  
  Run it yourself
&lt;/h3&gt;

&lt;p&gt;The weights are Apache 2.0:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;thx01
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;thx01&lt;/span&gt;
&lt;span class="n"&gt;agent&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;thx01&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;load&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;doofz/THX-01&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;agent&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;decide&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;My card was charged twice&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;team&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;choice&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;question&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Which team?&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
             &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;criteria&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;billing&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;billing&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;tech&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;technical&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;}}})&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It runs on a laptop CPU in a fraction of a second.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Model: &lt;a href="https://huggingface.co/doofz/THX-01" rel="noopener noreferrer"&gt;https://huggingface.co/doofz/THX-01&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Live demo: &lt;a href="https://huggingface.co/spaces/doofz/THX-01-demo" rel="noopener noreferrer"&gt;https://huggingface.co/spaces/doofz/THX-01-demo&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>opensource</category>
      <category>api</category>
      <category>machinelearning</category>
    </item>
    <item>
      <title>We Open-Sourced THX-01: A 322M AI Decision Model Small Enough to Run on a Smartphone</title>
      <dc:creator>FX-LgLL</dc:creator>
      <pubDate>Thu, 08 Oct 2026 21:26:50 +0000</pubDate>
      <link>https://dev.to/montydrief/we-open-sourced-thx-01-a-322m-ai-decision-model-small-enough-to-run-on-a-smartphone-27kg</link>
      <guid>https://dev.to/montydrief/we-open-sourced-thx-01-a-322m-ai-decision-model-small-enough-to-run-on-a-smartphone-27kg</guid>
      <description>&lt;p&gt;Today, we're introducing &lt;strong&gt;THX-01&lt;/strong&gt;, an open-source, non-autoregressive decision model developed by &lt;strong&gt;HAL-X AI&lt;/strong&gt; in Baku, Azerbaijan.&lt;/p&gt;

&lt;p&gt;With just &lt;strong&gt;322 million parameters&lt;/strong&gt;, THX-01 is lightweight enough for deployment on CPUs and potentially smartphones, without requiring expensive GPU infrastructure.&lt;/p&gt;

&lt;p&gt;Unlike traditional large language models, THX-01 doesn't generate text token by token. Instead, it processes inputs and returns structured decisions with calibrated confidence scores in a single forward pass.&lt;/p&gt;

&lt;h3&gt;
  
  
  Key Highlights
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;322M parameters&lt;/strong&gt;, built on mmBERT with a specialized decision head.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;~10 ms inference&lt;/strong&gt; per request on a GPU.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;CPU-compatible&lt;/strong&gt;, with a compact architecture suitable for mobile and edge deployment.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;98.4% accuracy&lt;/strong&gt; on our multilingual support-ticket classification benchmark, compared to 98.5% for Claude Sonnet 5.5.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;18 languages&lt;/strong&gt;, including Azerbaijani, English, Russian, and Turkish.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Calibrated confidence scores&lt;/strong&gt;, allowing applications to make decisions based on estimated uncertainty.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Apache 2.0 license&lt;/strong&gt;, with publicly available model weights.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Why We Built It
&lt;/h3&gt;

&lt;p&gt;Not every AI task requires a massive generative model.&lt;/p&gt;

&lt;p&gt;Classifying messages, routing customer requests, detecting spam, extracting document values, and making structured decisions can often be handled by significantly smaller models.&lt;/p&gt;

&lt;p&gt;THX-01 was designed specifically for these workloads, prioritizing speed, efficiency, and reliability.&lt;/p&gt;

&lt;p&gt;Our goal is to make intelligent decision-making accessible across environments, from production servers to resource-constrained edge devices.&lt;/p&gt;

&lt;h3&gt;
  
  
  Get Started
&lt;/h3&gt;

&lt;p&gt;Install THX-01 using pip:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;thx01
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Load the model:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;thx01&lt;/span&gt;

&lt;span class="n"&gt;agent&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;thx01&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;load&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;doofz/THX-01&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Open Source
&lt;/h3&gt;

&lt;p&gt;THX-01 is available under the Apache 2.0 license.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Model and documentation:&lt;/strong&gt; &lt;a href="https://huggingface.co/doofz/THX-01" rel="noopener noreferrer"&gt;https://huggingface.co/doofz/THX-01&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Built by HAL-X AI, Baku, Azerbaijan.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;We believe the future of AI isn't only about building larger models. It's also about building smaller, specialized models that make powerful capabilities accessible to everyone.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>programming</category>
      <category>opensource</category>
      <category>news</category>
    </item>
    <item>
      <title>We open-sourced THX-01: a 322M decision model that matches Sonnet 5.5 on ticket classification at ~10 ms (Apache 2.0, pip install thx01)</title>
      <dc:creator>FX-LgLL</dc:creator>
      <pubDate>Thu, 08 Oct 2026 18:24:29 +0000</pubDate>
      <link>https://dev.to/montydrief/we-open-sourced-thx-01-a-322m-decision-model-that-matches-sonnet-55-on-ticket-classification-at-5d53</link>
      <guid>https://dev.to/montydrief/we-open-sourced-thx-01-a-322m-decision-model-that-matches-sonnet-55-on-ticket-classification-at-5d53</guid>
      <description>&lt;p&gt;Today we're releasing THX-01, a model for the boring decisions that&lt;br&gt;
agents and backends make thousands of times a day: route this ticket, is this spam, which team, what's the invoice total,&lt;br&gt;
which sentence supports that.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What it is&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;322M parameters (mmBERT-base encoder + 15M decision head), Apache 2.0&lt;/li&gt;
&lt;li&gt;Non-autoregressive: you send a state plus typed questions, and it returns a probability for every option in one forward
pass. No generation, no JSON parsing, no retries.&lt;/li&gt;
&lt;li&gt;Question types: &lt;code&gt;choice&lt;/code&gt;, &lt;code&gt;noul&lt;/code&gt; (yes/no), &lt;code&gt;score&lt;/code&gt;, plus &lt;code&gt;number&lt;/code&gt; (pulls a value stated in the document, or &lt;code&gt;null&lt;/code&gt;),
&lt;code&gt;excerpt&lt;/code&gt; (verbatim span with offsets) and &lt;code&gt;"cite": true&lt;/code&gt; on any question&lt;/li&gt;
&lt;li&gt;~10 ms per request on one GPU, ~3,000 decisions/s batched, and it runs on CPU&lt;/li&gt;
&lt;li&gt;Speaks the &lt;code&gt;/v1/systemone&lt;/code&gt; wire format, so it is a drop-in for existing TypeSafe-style service layers&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Numbers (our support-ticket benchmark, 2,843 tickets in az/ru/en/tr, 15 categories)&lt;/strong&gt;&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;model&lt;/th&gt;
&lt;th&gt;avg acc&lt;/th&gt;
&lt;th&gt;latency&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;THX-01 (322M)&lt;/td&gt;
&lt;td&gt;98.4&lt;/td&gt;
&lt;td&gt;~10 ms&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Claude Sonnet 5.5&lt;/td&gt;
&lt;td&gt;98.5&lt;/td&gt;
&lt;td&gt;1.5 s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Wahoo 1.5&lt;/td&gt;
&lt;td&gt;97.8&lt;/td&gt;
&lt;td&gt;145 ms&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;TypeSafe Jev 1.13&lt;/td&gt;
&lt;td&gt;97.4&lt;/td&gt;
&lt;td&gt;331 ms&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Kev-4B&lt;/td&gt;
&lt;td&gt;92.8&lt;/td&gt;
&lt;td&gt;830 ms&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Because it's trained with a strictly proper scoring-rule reward, the probabilities are calibrated (ECE 0.003). With a single&lt;br&gt;
confidence threshold it handled ~75% of tickets fully automatically with zero errors on all four test sets.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;pip&lt;/span&gt; &lt;span class="n"&gt;install&lt;/span&gt; &lt;span class="n"&gt;thx01&lt;/span&gt;

&lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;thx01&lt;/span&gt;
&lt;span class="n"&gt;agent&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;thx01&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;load&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;doofz/THX-01&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;agent&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;decide&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;My card was charged twice for one order&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;team&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;choice&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;question&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Which team handles this?&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
             &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;criteria&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;billing&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;billing&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;tech&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;technical&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;sales&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;sales&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;}}})&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Model + docs: &lt;a href="https://huggingface.co/doofz/THX-01" rel="noopener noreferrer"&gt;https://huggingface.co/doofz/THX-01&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>opensource</category>
      <category>machinelearning</category>
      <category>llm</category>
    </item>
    <item>
      <title>SIMURG: A Guard That Catches LLM Hallucinations Mid-Stream and Heals Them</title>
      <dc:creator>FX-LgLL</dc:creator>
      <pubDate>Tue, 22 Sep 2026 12:01:37 +0000</pubDate>
      <link>https://dev.to/montydrief/simurg-a-guard-that-catches-llm-hallucinations-mid-stream-and-heals-them-252j</link>
      <guid>https://dev.to/montydrief/simurg-a-guard-that-catches-llm-hallucinations-mid-stream-and-heals-them-252j</guid>
      <description>&lt;p&gt;Your LLM looks fine in the demo. Then it ships.&lt;/p&gt;

&lt;p&gt;Somewhere between "it works on my machine" and "a user saw this", three failure modes show up in production streams:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;the model falls into a &lt;strong&gt;repetition loop&lt;/strong&gt; ("...and as we have seen, and as we have seen, and as we have seen..."),&lt;/li&gt;
&lt;li&gt;it &lt;strong&gt;drifts into another language&lt;/strong&gt; mid-answer,&lt;/li&gt;
&lt;li&gt;it &lt;strong&gt;collapses into structural garbage&lt;/strong&gt; — table echoes, &lt;code&gt;#REF!&lt;/code&gt; noise, token-level static.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;These are not model-intelligence failures. They are decoding failures, and they are detectable &lt;em&gt;while the stream is still open&lt;/em&gt;. That is the entire premise of &lt;a href="https://github.com/doofzoff/SIMURG" rel="noopener noreferrer"&gt;SIMURG&lt;/a&gt;: a streaming integrity monitor that watches every token as it arrives and raises a calibrated alarm within roughly 600 characters of the corruption onset — before the user has read the garbage.&lt;/p&gt;

&lt;p&gt;This week we shipped &lt;strong&gt;v1.0.4, "Catch The Heals"&lt;/strong&gt;, and it adds two things that change the game: the guard now &lt;em&gt;repairs&lt;/em&gt; the answer instead of discarding it, and it has a small trained deep-learning detector on board.&lt;/p&gt;

&lt;h2&gt;
  
  
  The problem with detecting corruption after the fact
&lt;/h2&gt;

&lt;p&gt;Post-hoc linters and LLM-as-judge pipelines all share one fatal property: they only see the full answer. By then the user has the garbage. Perplexity thresholds need logprob access most serving stacks do not expose. SIMURG's five-detector numpy ensemble works differently — it consumes the stream incrementally, scores every checkpoint, and can abort mid-generation.&lt;/p&gt;

&lt;p&gt;But aborting alone is expensive: a blind full retry means paying the model's wall-clock twice.&lt;/p&gt;

&lt;h2&gt;
  
  
  Self-Heal: diagnose, trim, continue, stitch
&lt;/h2&gt;

&lt;p&gt;The new repair ladder turns a corrupt abort into a targeted continuation:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Diagnose&lt;/strong&gt; — the fired corruption class is classified (repetition collapse, cross-lingual drift, regurgitation, structural breakdown).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Trim&lt;/strong&gt; — a periodic-loop onset detector (self-similarity plus exact-mismatch run scan) localizes the exact loop start, and the released prefix is trimmed to its &lt;em&gt;verified-clean&lt;/em&gt; boundary. The degenerate loop never enters the repair prompt.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Continue&lt;/strong&gt; — a pathology-specific instruction goes out at slightly warmer sampling: "continue from here, never repeat" for loops, "stay in the original language" for drift.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Stitch&lt;/strong&gt; — the continuation is guarded by a fresh sentinel; prefix plus verified tail are stitched into one answer, and &lt;code&gt;verify_final&lt;/code&gt; re-checks the whole thing. Zero-leak guarantee preserved.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The result: one generation's wall-clock instead of two, with &lt;code&gt;result.healed = True&lt;/code&gt; on the response object.&lt;/p&gt;

&lt;p&gt;In live A/B testing against a real endpoint (&lt;code&gt;wahoo-1.5-preview&lt;/code&gt; via VLLM) under high repetition pressure, the heal ladder scored &lt;strong&gt;5/5 clean outcomes vs 3/5&lt;/strong&gt; with healing disabled. At extreme penalty settings both drop — that is an honest limit of any guard, and it is documented as such.&lt;/p&gt;

&lt;h2&gt;
  
  
  SIMURG Pulse: the deep tier
&lt;/h2&gt;

&lt;p&gt;Statistics are robust, but they compress away &lt;em&gt;sequential&lt;/em&gt; structure: the exact phase of a repetition loop, the cadence of script switches, the texture of structural garbage. So v1.0.4 ships a sixth, learned view.&lt;/p&gt;

&lt;p&gt;Pulse is a 2-layer streaming transformer — 345K parameters, a 1.3 MB safetensors file. It reads the last ~600 characters of the stream as trigram-hash tokens (blake2b into 4096 buckets), runs a forward pass in about &lt;strong&gt;4 ms on Apple Silicon&lt;/strong&gt;, and emits a calibrated corruption probability via two anchors stored in the checkpoint metadata.&lt;/p&gt;

&lt;p&gt;Training protocol:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;40 clean live answers sampled from the guarded endpoint itself,&lt;/li&gt;
&lt;li&gt;240 synthetic corruptions (CorruptBench: loops, drift, table echo, structural garbage),&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;onset-aware labels&lt;/strong&gt; — a window counts as corrupt only once its right edge is 300+ characters past the true onset, so the model never learns to flag clean prefixes,&lt;/li&gt;
&lt;li&gt;BCE-with-logits with inverse-frequency class weighting, AdamW, 8 epochs, CPU-only training in seconds.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Held-out AUROC at training time: &lt;strong&gt;0.925&lt;/strong&gt;. On the release-time evaluation split (292 onset-aware windows) the ROC lands at 0.826 — small-clean-set variance, reported with the figure so nobody has to guess.&lt;/p&gt;

&lt;p&gt;Per-class response on single windows:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;input&lt;/th&gt;
&lt;th&gt;pulse probability&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;clean prose&lt;/td&gt;
&lt;td&gt;0.002&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;repetition loop&lt;/td&gt;
&lt;td&gt;1.000&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;cross-lingual drift&lt;/td&gt;
&lt;td&gt;1.000&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;table echo&lt;/td&gt;
&lt;td&gt;1.000&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;structural garbage&lt;/td&gt;
&lt;td&gt;1.000&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The contract is graceful by design: without &lt;code&gt;torch&lt;/code&gt; installed, the numpy-only core behaves exactly as before. &lt;code&gt;pip install "simurg[deep]"&lt;/code&gt; and the detector joins the ensemble automatically.&lt;/p&gt;

&lt;h2&gt;
  
  
  Train it on YOUR hallucinations
&lt;/h2&gt;

&lt;p&gt;This is the part we are most excited about. If your workload has a characteristic failure mode — fabricated citations, number drift, prompt echo, domain-specific garbage — the detector can be retrained against your endpoint in one command:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;export &lt;/span&gt;&lt;span class="nv"&gt;SIMURG_LIVE_URL&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;http://your-host:port/v1/chat/completions
&lt;span class="nb"&gt;export &lt;/span&gt;&lt;span class="nv"&gt;SIMURG_LIVE_MODEL&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;your-model-name

python &lt;span class="nt"&gt;-m&lt;/span&gt; simurg.deep.train_pulse &lt;span class="nt"&gt;--clean&lt;/span&gt; 40 &lt;span class="nt"&gt;--corrupt&lt;/span&gt; 240 &lt;span class="nt"&gt;--epochs&lt;/span&gt; 8 &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--out&lt;/span&gt; ./simurg_pulse.safetensors
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The trainer prints held-out AUROC and calibration anchors before saving. Point &lt;code&gt;SIMURG_PULSE_WEIGHTS&lt;/code&gt; at the new checkpoint and the ensemble picks it up on the next process start. Weights and the full model card are on Hugging Face: &lt;a href="https://huggingface.co/MergenAI/SIMURG" rel="noopener noreferrer"&gt;MergenAI/SIMURG&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Also in this release
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Free web search for agents&lt;/strong&gt; (TinyFish) — grounded answers out of the box, bundled key included.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Guard dashboard&lt;/strong&gt; — local UI for live guarding, session replay, and post-hoc analysis of pasted text.&lt;/li&gt;
&lt;li&gt;Paper: &lt;a href="https://ssrn.com/abstract=7451269" rel="noopener noreferrer"&gt;SIMURG: Zero-Leak Online Detection of LLM Decoding Corruption in Production Streams&lt;/a&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Links
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Code: &lt;a href="https://github.com/doofzoff/SIMURG" rel="noopener noreferrer"&gt;https://github.com/doofzoff/SIMURG&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Release: &lt;a href="https://github.com/doofzoff/SIMURG/releases/tag/v1.0.4" rel="noopener noreferrer"&gt;https://github.com/doofzoff/SIMURG/releases/tag/v1.0.4&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Weights: &lt;a href="https://huggingface.co/MergenAI/SIMURG" rel="noopener noreferrer"&gt;https://huggingface.co/MergenAI/SIMURG&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;PyPI: &lt;a href="https://pypi.org/project/simurg/1.0.4/" rel="noopener noreferrer"&gt;https://pypi.org/project/simurg/1.0.4/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Paper: &lt;a href="https://ssrn.com/abstract=7451269" rel="noopener noreferrer"&gt;https://ssrn.com/abstract=7451269&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;License: Apache-2.0.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>opensource</category>
      <category>automation</category>
      <category>machinelearning</category>
    </item>
    <item>
      <title>Free Unlimited Web Search for AI Agents</title>
      <dc:creator>FX-LgLL</dc:creator>
      <pubDate>Wed, 02 Sep 2026 18:59:51 +0000</pubDate>
      <link>https://dev.to/montydrief/free-unlimited-web-search-for-ai-agents-1p56</link>
      <guid>https://dev.to/montydrief/free-unlimited-web-search-for-ai-agents-1p56</guid>
      <description>&lt;p&gt;AI agents need access to the web.&lt;/p&gt;

&lt;p&gt;The problem is that web search APIs can be expensive, limited, or complicated to integrate.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;SIMURG Search&lt;/strong&gt; gives AI agents a simple way to access web search for free.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is SIMURG?
&lt;/h2&gt;

&lt;p&gt;SIMURG is an open-source protection layer for LLMs.&lt;/p&gt;

&lt;p&gt;It monitors model generation and detects when the output starts going wrong, such as repetition loops, language drift, garbage output, or other decoding failures. When corruption is detected, SIMURG can stop the generation and retry it before the bad output reaches the user.&lt;/p&gt;

&lt;p&gt;SIMURG also includes a web search layer that allows agents to retrieve real information from the internet and use it for grounding before answering.&lt;/p&gt;

&lt;h2&gt;
  
  
  SIMURG Search
&lt;/h2&gt;

&lt;p&gt;With SIMURG Search, your AI agent can search the web, retrieve sources, and use the results to verify information instead of relying only on the model's memory.&lt;/p&gt;

&lt;p&gt;It is designed to be simple:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;simurg
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then you can use the built-in web search directly from your agent.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why use it?
&lt;/h2&gt;

&lt;p&gt;Free web search for AI agents&lt;/p&gt;

&lt;p&gt;Open source&lt;/p&gt;

&lt;p&gt;Simple installation&lt;/p&gt;

&lt;p&gt;Real-time web information&lt;/p&gt;

&lt;p&gt;Useful for local, quantized and self-hosted models&lt;/p&gt;

&lt;p&gt;SIMURG combines generation protection with web grounding, giving AI agents a simple way to become more reliable.&lt;/p&gt;

&lt;h2&gt;
  
  
  Get Started
&lt;/h2&gt;

&lt;p&gt;Install SIMURG:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;simurg
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;GitHub repository:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/doofzoff/SIMURG" rel="noopener noreferrer"&gt;https://github.com/doofzoff/SIMURG&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;If your AI agent needs web access without adding another complicated infrastructure layer, give SIMURG Search a try.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>opensource</category>
      <category>automation</category>
      <category>machinelearning</category>
    </item>
    <item>
      <title>I found a completely free web search for AI agents.</title>
      <dc:creator>FX-LgLL</dc:creator>
      <pubDate>Sat, 29 Aug 2026 19:21:02 +0000</pubDate>
      <link>https://dev.to/montydrief/i-found-a-completely-free-web-search-for-ai-agents-5an4</link>
      <guid>https://dev.to/montydrief/i-found-a-completely-free-web-search-for-ai-agents-5an4</guid>
      <description>&lt;p&gt;A week ago my agent confidently described a treaty that does not exist. It even had a date.&lt;/p&gt;

&lt;p&gt;The fix is simple: before the model commits to an answer, make it check the web. &lt;br&gt;
I contributed this to SIMURG&lt;br&gt;
(&lt;a href="https://github.com/doofzoff/SIMURG" rel="noopener noreferrer"&gt;https://github.com/doofzoff/SIMURG&lt;/a&gt;) and it just shipped as v1.0.3.&lt;/p&gt;

&lt;h2&gt;
  
  
  Free means free
&lt;/h2&gt;

&lt;p&gt;SIMURG web search is a web layer for AI agents powered by TinyFish. 30&lt;br&gt;
requests per minute, $0, no card, no wallet. The package ships a bundled free&lt;br&gt;
key, so it works out of the box:&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;pip install simurg
python3 -m simurg.websearch "when was the Y2K bug" --json
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;
&lt;h2&gt;
  
  
  The part I care about: ground()
&lt;/h2&gt;

&lt;p&gt;Search results are context. Verdicts are decisions:&lt;/p&gt;
&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;from simurg import websearch
check = websearch.ground("Y2K bug")
# check["verdict"]: "attested" | "thin" | "no_record"
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;attested: feed check["evidence"] into your model's context. thin: treat with&lt;br&gt;
caution. no_record: the subject is probably fabricated, abstain instead of&lt;br&gt;
asserting.&lt;/p&gt;

&lt;p&gt;One honest detail: "Zorbachian treaty of 1874" returns real 1874 treaties from&lt;br&gt;
any search engine. So ground() requires the subject itself to be echoed in the&lt;br&gt;
evidence, generic hits downgrade to thin.&lt;/p&gt;

&lt;p&gt;Stdlib only, zero new dependencies, and it degrades instead of crashing. In&lt;br&gt;
our internal checks on factual prompts, hallucinations dropped to the 1-5%&lt;br&gt;
range.&lt;/p&gt;

&lt;p&gt;Code, examples and a 13 page paper: &lt;a href="https://github.com/doofzoff/SIMURG" rel="noopener noreferrer"&gt;https://github.com/doofzoff/SIMURG&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Give your agent an internet. It costs nothing.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>programming</category>
      <category>opensource</category>
      <category>machinelearning</category>
    </item>
  </channel>
</rss>
