<?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: Yamis</title>
    <description>The latest articles on DEV Community by Yamis (@yamis).</description>
    <link>https://dev.to/yamis</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%2F4252685%2F6835f583-4a70-4866-ad1a-054f6a67db43.png</url>
      <title>DEV Community: Yamis</title>
      <link>https://dev.to/yamis</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/yamis"/>
    <language>en</language>
    <item>
      <title>Running your AI at home and sharing it safely – lessons from building Yamis</title>
      <dc:creator>Yamis</dc:creator>
      <pubDate>Fri, 20 Nov 2026 12:15:43 +0000</pubDate>
      <link>https://dev.to/yamis/running-your-ai-at-home-and-sharing-it-safely-lessons-from-building-yamis-4kfe</link>
      <guid>https://dev.to/yamis/running-your-ai-at-home-and-sharing-it-safely-lessons-from-building-yamis-4kfe</guid>
      <description>&lt;p&gt;For the last few months I've been building &lt;strong&gt;Yamis&lt;/strong&gt;, an AI assistant whose model runs on the user's own computer. No subscription, no prompts and files sitting on someone else's AI server. Here is what I learned along the way – including the parts that were much harder than I expected.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why run the AI at home?
&lt;/h2&gt;

&lt;p&gt;Three reasons kept coming up:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Cost.&lt;/strong&gt; Lots of people pay a monthly AI subscription while a perfectly capable PC sits idle next to them.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Privacy.&lt;/strong&gt; Mails, documents, half-finished ideas – people are rightly uneasy about sending all of that to a company where nobody really knows what happens with it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The hardware is already there.&lt;/strong&gt; Many gaming PCs and workstations have enough graphics memory to run a good open model. Even without a big GPU, mixture-of-experts models run surprisingly well on CPU and RAM.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The problem: setting this up is still a job for experts. Which model fits my machine? Which quantization? How do I reach it from my phone without opening my router to the internet? That is the gap Yamis tries to close.&lt;/p&gt;

&lt;h2&gt;
  
  
  The hard parts
&lt;/h2&gt;

&lt;h3&gt;
  
  
  1. Picking a model that actually fits
&lt;/h3&gt;

&lt;p&gt;"Use the biggest model that fits in VRAM" sounds simple, but it isn't. The weights are only part of it – the context (KV cache) needs memory too, and with long contexts that can be several gigabytes. Mixture-of-experts models can keep their experts in system RAM and still be fast, because only a small part of the model is active per token.&lt;/p&gt;

&lt;p&gt;So the desktop helper (&lt;strong&gt;Mimir&lt;/strong&gt;) measures the machine first: GPU, VRAM, RAM, CPU features. It then calculates which models and quantizations fit with a safety margin and – this turned out to matter most – measures real tokens per second instead of trusting a table. A model that technically fits but answers at two tokens per second is not a good recommendation.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. No open port
&lt;/h3&gt;

&lt;p&gt;I did not want users to forward ports or expose anything on their home network. Instead Mimir keeps an &lt;strong&gt;outbound&lt;/strong&gt; connection to Yamis open and asks for work. When a request comes in, Yamis hands it over this connection, Mimir runs it on the local llama.cpp server and streams the answer back. Nothing listens on the user's router, and router credentials (used for Wake-on-LAN, for example) never leave the device.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Small models make things up
&lt;/h3&gt;

&lt;p&gt;Local models are good, but smaller ones are more confident than they should be. They "remember" files they never opened and report tests as green that never ran. So Yamis doesn't just trust the answer:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;a plan comes first, and changes on your computer only happen after you say yes;&lt;/li&gt;
&lt;li&gt;claims like "the build is green" need evidence from an actual run;&lt;/li&gt;
&lt;li&gt;a second pass reviews the changes before the result is handed over – including a check for logic that was silently removed.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This part is never finished, but it made the difference between a demo and something I'd let touch my own machines.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sharing without risk
&lt;/h2&gt;

&lt;p&gt;Once your AI runs at home, the obvious next question is: can my partner, my parents or my team use it too? Yamis lets you &lt;strong&gt;share your AI&lt;/strong&gt; with other accounts:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;a daily limit per person,&lt;/li&gt;
&lt;li&gt;revocation with one click,&lt;/li&gt;
&lt;li&gt;and the people you share with never see your machine – they only see the answers.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;You can also work together live: several people in one chat with the AI (we call it &lt;em&gt;Tandem&lt;/em&gt;).&lt;/p&gt;

&lt;h2&gt;
  
  
  What else is in there
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Chat in the browser, the phone app or a terminal CLI (also over SSH on servers).&lt;/li&gt;
&lt;li&gt;Agentic tasks you can plan and schedule.&lt;/li&gt;
&lt;li&gt;Your own knowledge (mails, files) – optionally stored encrypted at home.&lt;/li&gt;
&lt;li&gt;An OpenAI-compatible API, so your own tools can use your home AI.&lt;/li&gt;
&lt;li&gt;9 languages.&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;The closed beta starts on &lt;strong&gt;December 1st&lt;/strong&gt; and is free. Beta testers can tell the AI what they're missing with &lt;code&gt;/support&lt;/code&gt;, and sensible ideas get built in.&lt;/p&gt;

&lt;p&gt;Waiting list: &lt;a href="https://yamis.it" rel="noopener noreferrer"&gt;yamis.it&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What do you think about local AI?
&lt;/h2&gt;

&lt;p&gt;I'm curious about your take: would you run your AI at home – and maybe share it with friends or family – or is a cloud subscription simply more convenient for you? What would stop you from switching? Let me know in the comments.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>privacy</category>
      <category>showdev</category>
    </item>
  </channel>
</rss>
