DEV Community: AI OpenFree

Sovereign security AI: Mythos-class defense on your own GPUs (Chitos)

AI OpenFree — Sun, 02 Aug 2026 07:40:35 +0000

Sovereign security AI: Mythos-class defense on your own GPUs (Chitos)

Cloud attack-security AI (e.g. Anthropic Mythos) is powerful but cloud-only — your code leaves your building to use it. For defense/gov/finance/healthcare, that is a non-starter.

Chitos is a sovereign, on-prem autonomous security-AI harness: find it before attackers do, then prove it — with no data egress.

Two engines

Deterministic static analysis — reproducible detection (injections, deserialization, SSRF/XXE/XSS, credential exposure, memory safety, auth, logic), CWE-mapped, no AI needed (0 cost, 0 egress).
AI reasoning layer — verifies + explains findings, reasons kill chains. => dual evidence vs one black-box verdict.

Trust features

Metacognitive self-check: before asserting a vuln, it asks false positive? reproduces? real source to sink flow? -> structurally fewer false alarms.
CodeGraph grounding: maps definitions/deps/risky call sites to localize analysis.
3 modes: Scan / Research (CVE web lookup) / Exploit (kill chain on approved target). Code snippet or URL input.
Default: DeepSeek V4 Flash (304B MoE, MIT) on your infra; Claude via your key optional.

Honest scope

Mythos is stronger at raw discovery — Chitos does not claim otherwise. Chitos wins on sovereignty, transparency, reproducibility, regulatory fit, cost. Mythos-class defense, on your own GPUs, no egress.

Free, no signup: https://chitos.vidraft.net

Build quantum circuits with natural language: VIDRAFT free GPU qubit simulator

AI OpenFree — Sun, 02 Aug 2026 03:48:40 +0000

Build quantum circuits with natural language: VIDRAFT free GPU qubit simulator (33 qubits, H100)

VIDRAFT opened quantum.1street.ai — a browser quantum-circuit simulator with an AI copilot: describe what you want in plain language and it builds the circuit.

What it does

Talk to build: say “make it 0 and 1 fifty-fifty like a coin” -> the AI copilot generates the circuit.
One-click examples: entanglement, Grover search, QFT, molecular energy (VQE), error correction.
Manual mode: drop H/X/Y/Z/S/T/RY/CX gates on qubit wires.
Visuals: measurement probabilities, per-qubit Bloch spheres, entanglement bars.

Capacity (measured)

33 qubits exact precision on 1x H100 (34 on 4x H100 = 17.2B states).
100,000 qubits for Clifford (flip/entangle) circuits via a math shortcut.
Thousands for low-entanglement circuits (MPS).

Honest scope

It is a GPU simulator, not a real QPU — for education, intuition and prototyping, not quantum speedups. KO/EN supported with auto browser-language detection.

🔗 https://quantum.1street.ai

Follow the Build Logs: VIDRAFT's Blog and the Founder's Brunch Essays

AI OpenFree — Sat, 01 Aug 2026 06:42:14 +0000

Follow the Build Logs: VIDRAFT's Blog and the Founder's Brunch Essays

Notes from inside the build

The VIDRAFT blog is where the work gets explained in plain language. It carries two kinds of writing: editorial notes on where AI is heading, and engineering notes on how we actually built and shipped things. It is aimed at developers, so it favors specifics over announcements -- what we tried, what broke, and what we changed as a result.

Two streams to follow

There are two places to read us.

The blog itself holds engineering write-ups and editorial posts tied to the products: model releases, benchmark methodology, deployment lessons, and the occasional retrospective when something did not go to plan.

Separately, the founder writes long-form essays on Kakao Brunch under the handle @seawolf. Brunch is a Korean long-form publishing platform, and the essays there take a wider lens -- strategy, the Korean AI landscape, and the reasoning behind the bets a small lab makes. If the blog is "here is how we built it," the Brunch essays are "here is why we chose to build it this way."

Why we publish post-mortems

Most company blogs are highlight reels. We would rather publish the honest post-mortem: the merge that regressed, the benchmark that turned out to be measuring the wrong thing, the deployment that failed in a way we did not anticipate. For a lab whose entire pitch is method over scale, the method only earns trust if the failures sit on the record next to the wins. A clean narrative is easy to write and easy to distrust. A post-mortem that names the mistake is harder to fake.

For developers this is also the most useful content we can produce. A write-up of what broke in a real deployment saves you from hitting the same wall. That is the exchange we are after: we share the build logs, you get to skip our mistakes, and the feedback makes the next model better.

What to expect

Cadence over polish. Posts land when there is something real to report -- a release, a result, a lesson -- rather than on a marketing calendar. Expect tables, benchmark numbers with their scope attached, and code where code makes the point better than prose.

Honest scope

This is editorial and engineering writing, not a support channel or a spec sheet. Opinions in the posts are ours and may age badly as the field moves -- that is the nature of writing in public about a fast-moving area. The Brunch essays are the founder's personal long-form writing and reflect an individual viewpoint, not formal company positions or commitments. Nothing in a blog post is a product guarantee; when we describe something as in progress, it is in progress. Follow along for the reasoning and the post-mortems, and hold us to what we ship.

More: https://vidraft.net

VIDRAFT: A Capital-Efficient Challenger, Method Over Scale

AI OpenFree — Sat, 01 Aug 2026 05:42:15 +0000

VIDRAFT: A Capital-Efficient Challenger, Method Over Scale

A small Seoul lab, on purpose

VIDRAFT Inc. was founded in 2024 in Seoul. The entire thesis is method over scale: build foundation models, on-device models, and AI-safety benchmarks with a fraction of the compute that larger labs treat as mandatory, and let the results carry the argument.

The numbers that matter

Fact	Detail
Founded	2024, Seoul
Compute	~24 GPUs (B200 x16 + H200 x8)
Grants	NIPA GPU grants -- 2025 AI computing, 2026 Blackwell
Downloads	~1.18M cumulative (HF-verified, 2026-07)
Leaderboard	#1 on the K-AI leaderboard (leaderboard scope)
Polaris	15 first places
HF orgs	VIDraft, FINAL-Bench

About 24 GPUs is a small cluster by frontier-lab standards. The NIPA GPU grants -- 2025 AI computing, and 2026 Blackwell -- are how a company this size gets access to modern hardware at all, and they are a big part of why the capital-efficient approach is a necessity rather than just a slogan.

What we build

Foundation models via evolutionary merging. On-device models small enough to run outside a data center. AI-safety benchmarks, published under the FINAL-Bench org. The models and their community copies ship on Hugging Face under the VIDraft org, where cumulative downloads reached roughly 1.18M as of July 2026. GINIGEN is the surrounding ecosystem those products live in.

The through-line is efficiency as a design constraint, not an afterthought. When you cannot outspend the field, you have to out-method it -- reuse checkpoints instead of pretraining from zero, optimize for hardware you actually have, and measure everything so you know which trades paid off. A ~24-GPU cluster is not a limitation we apologize for; it is the premise the whole approach is built to respect.

Track record, scoped

We are #1 on the K-AI leaderboard, and we hold 15 Polaris first places. Both are real, and both are scoped: a leaderboard rank is rank on that leaderboard, against the field that submitted, under its rules. We state the scope every time because a challenger's credibility depends on not overselling.

Honest scope

Read every claim here with its scope attached. "#1 on the K-AI leaderboard" means first on that specific leaderboard, not a global title. The ~1.18M download figure is cumulative across our AI models plus community copies and quantizations -- a reach signal, not a count of paying customers or unique users. Founded-2024 means we are early: small team, small cluster, short track record. The NIPA grants are compute grants, not equity or endorsement. We publish these numbers because a challenger earns trust by being precise about what it has and has not done, not by rounding up.

More: https://vidraft.net

Sovereign AX: Frontier-Class AI With No Data Egress

AI OpenFree — Sat, 01 Aug 2026 04:42:13 +0000

Sovereign AX: Frontier-Class AI With No Data Egress

Run the model where your data already lives

The core promise of VIDRAFT's Solutions and Services line is simple: run frontier-class AI entirely on your own infrastructure, with no data egress. For banks, hospitals, and government, the blocker to adopting AI is rarely capability -- it is that the data cannot leave the building. Sovereign AX (AI transformation) is built around that constraint instead of against it.

What sovereign AX includes

On-prem foundation models. We deploy our own models inside your environment. Weights, inference, and logs stay on hardware you control. Nothing is sent to an external API.
A sovereign VLM. A vision-language model for document, image, and multimodal understanding, currently in collaboration discussions -- for example with DeepBrain AI. The aim is frontier-class multimodal capability that never phones home.
On-device deployment. Beyond the data center, our on-device model work pushes inference to edge hardware -- kiosks, workstations, embedded devices -- where connectivity or policy rules out the cloud entirely.
A full service catalog. Fine-tuning on your data, evaluation, and integration, delivered as engagements rather than as a black box.

Concern	Cloud AI API	Sovereign AX
Where data goes	To a third-party endpoint	Stays on your infrastructure
Model location	Vendor cloud	Your data center or device
Auditability	Vendor-controlled	You hold the logs

Why a small team can offer this

Our whole model-building method is capital-efficient by design -- evolutionary merging and on-device optimization aim for frontier-class behavior without frontier-scale compute. That same efficiency is what makes on-prem deployment realistic: a model tuned to run on a constrained footprint is exactly the model an enterprise can host itself, without renting a hyperscaler's fleet.

Where it fits

Sovereign AX makes sense when the data is the whole problem: regulated records, classified material, proprietary corpora that legal will never let touch a public endpoint. If your workload is casual and your data is already public, a hosted API is probably cheaper. We would rather tell you that than sell you an on-prem deployment you do not need.

Honest scope

This is a services-and-deployment offering, not a shrink-wrapped product you download and run in an afternoon. "Frontier-class" describes the target and our benchmark and leaderboard results, not a guarantee that a given model matches the largest closed labs on every task -- capability depends on the workload, the domain, and the fine-tuning data you provide. The sovereign VLM is in collaboration discussions; naming DeepBrain AI describes an active conversation, not a shipped joint product or an endorsement. On-prem deployment moves real integration work to your side of the fence: hardware, security review, and tuning are part of the engagement. What we do commit to is the architecture -- no data egress -- and to being explicit about what is production-ready versus in progress.

More: https://vidraft.net

Four Preprints and a Patent Stack: What VIDRAFT Is Actually Publishing

AI OpenFree — Sat, 01 Aug 2026 04:40:23 +0000

Four Preprints and a Patent Stack: What VIDRAFT Is Actually Publishing

Research on a small-lab budget

VIDRAFT publishes. For a team competing with much larger labs, writing things down -- preprints and patent filings -- is how we turn "method over scale" into something checkable instead of rhetorical. Here is what is actually on the record, with identifiers so you can go read the primary sources.

Four preprints

ID	Topic
arXiv 2605.14386	Darwin: model merging
arXiv 2607.18340	Quantum
arXiv 2607.24830	A numerical realization related to the Riemann/Weil line
SSRN 6280258	AI metacognition

The Darwin paper is the backbone of how we build models: evolutionary merging, where you search over combinations of existing checkpoints to produce a stronger model instead of pretraining a new one from scratch. It is the most direct expression of the capital-efficient thesis -- compute spent on search rather than on raw pretraining.

The quantum preprint and the Riemann/Weil-line numerical work sit further from product and closer to basic research. The metacognition paper, posted on SSRN, studies whether a model can estimate its own reliability -- knowing when it does not know -- which feeds the AI-safety benchmark work we run under the FINAL-Bench org.

The patent portfolio

Alongside the papers there is a patent portfolio: 12 filings with 317 claims in total. Patents and preprints do different jobs. Preprints invite scrutiny and citation; filings stake defensible ground around the methods. A small company needs both -- open enough to be credible, protected enough to be fundable.

How to read all of this

Treat the identifiers as invitations, not trophies. Pull up arXiv 2605.14386 and check whether the merging method does what the abstract claims. Look at the metacognition setup on SSRN and decide whether the evaluation convinces you. That is the whole point of putting a number on the record: it lets a reader replace our summary with their own judgment.

We publish because a challenger earns credibility by exposing its reasoning, not by asking for trust. The same discipline that keeps our compute budget honest -- measure it, write it down, let others check -- applies to the research.

Honest scope

These are recent preprints. Citation counts and social traction are early-stage -- low, in plain terms -- because the work is new and, in some cases, narrow. A preprint is not peer-reviewed: an arXiv or SSRN ID means the work is public and timestamped, not that a community has validated it. The Riemann/Weil-line entry is a numerical realization, not a claimed proof of anything. Patent filings are filings -- claim counts describe scope on paper, not granted, litigated, or commercially proven rights. We list the identifiers precisely so you can judge for yourself, and if a result does not replicate, tell us. That feedback is why we publish.

More: https://vidraft.net

Two Ways to Simulate a Planet: EarthOS and NationalOS

AI OpenFree — Sat, 01 Aug 2026 02:42:16 +0000

Two Ways to Simulate a Planet: EarthOS and NationalOS

Simulation at two scales

VIDRAFT runs two large-scale simulation efforts that sit at opposite ends of the map. EarthOS models the physical planet from observation data. NationalOS models a human population as a swarm of AI agents. Both are bets that a small team can build simulation infrastructure usually associated with much larger labs -- method over scale.

EarthOS: an Earth-observation foundation model

EarthOS is an Earth-observation foundation model that absorbs the OlmoEarth approach: pretrain on satellite and remote-sensing data to produce embeddings that transfer to downstream geospatial tasks such as land cover, change detection, crop and vegetation analysis, and disaster response.

The design goal is a shared representation. Instead of training a bespoke model per sensor or per task, you learn one backbone over multi-modal Earth observation and adapt it with lightweight heads. That is the same transfer-learning logic that made language and vision foundation models useful, now applied to pixels that arrive from orbit.

License matters here. EarthOS ships under terms that exclude surveillance and policing use. Earth-observation models are dual-use by nature, and we would rather constrain the license up front than pretend the risk is not there.

NationalOS: a society twin

NationalOS is a different kind of simulation -- a society twin. Instead of physical pixels it models a population as millions of AI-agent NPCs, each an individual with its own attributes and behavior. The Korea build targets roughly 52 million agents, on the order of one per person in the country. A USA variant applies the same approach to a different population.

The point is emergent, population-level behavior. When you instantiate millions of interacting agents you can pose what-if questions about policy, information flow, and collective response that aggregate statistical models handle poorly. It is agent-based modeling pushed to national scale.

System	Models	Scale	Primary use
EarthOS	The physical planet from EO data	Foundation-model embeddings	Geospatial downstream tasks
NationalOS	A human population	~52M agents (Korea build)	Policy and society simulation

Why build both

The two share a worldview: complex systems are easier to reason about when you have a learned, queryable model of them rather than a spreadsheet. EarthOS gives you a substrate for the environment. NationalOS gives you a substrate for the people who live in it. In principle they compose -- environmental signals feeding a population model -- though that integration is future work, not a shipped feature.

Running both on a ~24-GPU footprint forces discipline. We cannot brute-force these problems, so the engineering leans on transfer learning, efficient agent representations, and reuse of pretrained backbones rather than training everything from scratch.

Honest scope

Both systems are research-stage. EarthOS is a foundation-model effort following a published approach, not a certified operational forecasting service; downstream accuracy depends entirely on the task and the fine-tuning data you bring. NationalOS is a simulation of behavior, not a prediction oracle -- agent-based models illustrate dynamics and stress-test assumptions, but their outputs are only as good as the behavioral priors encoded in the agents, and they should never be read as literal forecasts of what specific real people will do. Agent counts describe modeling scale, not validated demographic fidelity. We are sharing these as active builds and will report concrete evaluations as they mature.

More: https://vidraft.net

Physical AI: A 4B Vision-Language-Action Model and the Sim-to-Real Gap We Do Not Hide

AI OpenFree — Sat, 01 Aug 2026 01:42:15 +0000

Physical AI: A 4B Vision-Language-Action Model and the Sim-to-Real Gap We Do Not Hide

Robots need a different kind of model

A chatbot's output is text. A robot's output is torque, grasp, and motion - actions that fail in the physical world if the model misjudges geometry, timing, or contact. Physical AI is VIDRAFT's research line for that harder output space: vision-language-action (VLA) models that take pixels and instructions in, and emit actions out.

Darwin-4B-Robo

Darwin-4B-Robo is our VLA model. On the RoboCasa 24-task benchmark it ranked #1 (leaderboard scope) - a simulated manipulation suite covering 24 distinct household-style tasks. At 4B parameters it is deliberately small: embodied policies have to run on real robots with real latency budgets, so parameter efficiency is a feature, not a compromise.

The model builds on the David + pi-0 lineage - prior open work in the VLA space - rather than starting from scratch. Standing on that lineage is the honest way to move fast in robotics research, where reproducible baselines are scarce.

Why small models, and why simulation first

Embodied research runs on a tight loop: a policy that cannot execute inside a robot's control cycle is useless no matter how accurate. That is why Darwin-4B-Robo is small. Simulation benchmarks like RoboCasa let us iterate hundreds of times faster than on hardware and compare against other methods on equal footing - but we treat sim numbers as a leading indicator, not a finish line. The real test is always physical, and that is the harder, slower work ahead.

World models

The longer-term thread is world-model research: giving a policy an internal predictive model of how the environment responds to its actions, so it can plan rather than react. A VLA that only maps observation to action is reactive; one that can roll out "if I push here, this happens" is a step toward robust manipulation. This part is early.

Honest scope

The RoboCasa #1 result is scoped to that benchmark's leaderboard and its task set; RoboCasa is a simulation suite, so it measures simulated manipulation, and sim performance does not automatically transfer to physical robots (the sim-to-real gap is real and unsolved in general). Darwin-4B-Robo is embodied-AI research, not a shipping robotics product, and the world-model direction is research-stage. The David + pi-0 lineage is prior work we build on and credit, not our invention. Read this as a research report from an active line, not a product datasheet.

More: https://vidraft.net

QuantumOS: An Even-Mansour Key-Recovery Demo on Real Quantum Hardware (and What It Is Not)

AI OpenFree — Sat, 01 Aug 2026 00:42:14 +0000

QuantumOS: An Even-Mansour Key-Recovery Demo on Real Quantum Hardware (and What It Is Not)

Cryptanalysis on real quantum hardware

QuantumOS is VIDRAFT's quantum-computing research line. The headline experiment: a key-recovery demonstration against the Even-Mansour construction, run on real IBM quantum hardware at N=10 qubits. The work was referenced in SemiEngineering's weekly industry review (external reference).

What Even-Mansour is, and what we did

Even-Mansour is a minimalist block-cipher construction - a public permutation wrapped by two XOR'd secret keys. It is a favorite target for quantum cryptanalysis research because its structure is clean enough to study on the small, noisy quantum machines that exist today. We implemented a key-recovery attack and ran it end-to-end on actual quantum hardware, not just a simulator.

What this is NOT

Being blunt, because this is the area where quantum claims get oversold:

This is a small-scale demonstration at N=10 qubits. It is not a practical attack.
It does not break AES, RSA, elliptic-curve, or any cryptography deployed in the real world.
It is not a demonstration of quantum speedup at any useful scale. Running on real hardware proves the circuit executes and returns the expected structure under noise - it does not prove an advantage over classical machines.

Nothing you use is weaker because this experiment ran. The value is scientific: validating that these attacks can be expressed and executed on today's hardware, and building the tooling and intuition for when the machines get bigger.

Why run it on real hardware at all

Simulators are convenient and, for 10 qubits, arguably easier. Running on physical hardware forces you to confront noise, decoherence, and error rates that a clean simulator hides. That is the point: the interesting research question is not "does the math work" (it does, on paper) but "what survives on a real, noisy device, and what breaks." Answering that now is how you are ready when qubit counts and fidelity improve.

Sibling research lines

QuantumOS sits alongside two other VIDRAFT research efforts that share the same run-real-physics-on-emerging-hardware mindset:

PharmaOS - drug-discovery research.
MaterialsOS - materials research.

All three are research-stage.

Honest scope

The Even-Mansour result is a research-stage, small-scale (N=10) hardware demonstration. It is explicitly not a practical quantum attack, not a speedup claim, and not a break of any deployed cryptosystem - do not read it as "quantum computers can break encryption now," because that is false. The SemiEngineering mention is an industry-review reference, not a peer-reviewed endorsement of a cryptographic break. PharmaOS and MaterialsOS are likewise early-stage research lines, mentioned here for context, not as shipping products. Quantum hardware today is small and noisy; honest reporting of that constraint is part of the work.

More: https://vidraft.net

FINAL-Bench: A Diagnostic for Whether a Model Knows What It Does Not Know

AI OpenFree — Fri, 31 Jul 2026 23:42:16 +0000

FINAL-Bench: A Diagnostic for Whether a Model Knows What It Does Not Know

What FINAL-Bench measures

Most benchmarks ask "did the model get the right answer?" FINAL-Bench asks a harder, safety-relevant question: does the model know what it does not know? It is a metacognition and safety diagnostic - it probes calibration, hallucination-awareness, and whether a model refuses appropriately instead of confidently making things up.

Why calibration, not accuracy

A model that is wrong 20% of the time but knows which 20% is far safer to deploy than a model that is wrong 10% of the time and confident about all of it. In regulated settings - health, finance, law - the failure that hurts is confident wrongness. FINAL-Bench targets exactly that axis: the gap between what a model claims to know and what it actually knows.

Two signals we can share:

Single-pass self-doubt is surprisingly informative. A model's own hesitation, read from a single forward pass, separates correct from incorrect answers with AUROC around 0.90 (research-stage signal).
But a model's own confidence is a weak judge of itself. An external judge scores around AUROC 0.72, while the model grading its own confidence lands near 0.59. The lesson: self-reported confidence is not enough; independent evaluation beats introspection.

Mapping to regulation

Findings are mapped, at a high level, to the frameworks teams actually answer to - the EU AI Act, the NIST AI Risk Management Framework, and similar. The goal is to turn an abstract "is this model safe?" question into evidence a risk or compliance team can read: where the model over-claims, where it refuses well, where it needs a human gate.

The diagnostic is built from 117 items.

Who this is for

Teams shipping models into regulated or high-stakes settings, and the risk and compliance people who have to sign off. If you already track accuracy, FINAL-Bench adds the orthogonal axis you are probably missing: whether the model's confidence can be trusted at all.

Honest scope

The 117-item methodology - how items are constructed and how responses are scored - is patent-pending and under NDA. This article describes only the purpose (metacognition and safety), the calibration idea, and the regulatory mapping; it deliberately does not disclose item design or scoring internals. The AUROC figures (single-pass self-doubt ~0.90, external judge ~0.72, model self-confidence ~0.59) are research-stage signals from our own runs, not certified scores for any specific commercial model. FINAL-Bench is a diagnostic, not a certification: it produces evidence and risk mapping, it does not "pass" or "approve" a model for regulated use, and the law-mapping is an interpretive aid, not legal advice. Independent replication and third-party audit are the right next step, and we treat it that way.

More: https://vidraft.net

FINAL-Bench: A Diagnostic for Whether a Model Knows What It Does Not Know

AI OpenFree — Fri, 31 Jul 2026 18:41:22 +0000

FINAL-Bench: A Diagnostic for Whether a Model Knows What It Does Not Know

What FINAL-Bench measures

Why calibration, not accuracy

Two signals we can share:

Single-pass self-doubt is surprisingly informative. A model's own hesitation, read from a single forward pass, separates correct from incorrect answers with AUROC around 0.90 (research-stage signal).
But a model's own confidence is a weak judge of itself. An external judge scores around AUROC 0.72, while the model grading its own confidence lands near 0.59. The lesson: self-reported confidence is not enough; independent evaluation beats introspection.

Mapping to regulation

The diagnostic is built from 117 items.

Who this is for

Honest scope

More: https://vidraft.net

VICODE + CHITOS: Run a Full Coding Agent Without Your Code Leaving the Building

AI OpenFree — Fri, 31 Jul 2026 10:42:14 +0000

VICODE + CHITOS: Run a Full Coding Agent Without Your Code Leaving the Building

Run a full coding agent without your code leaving the building

VICODE is an on-prem coding agent: a self-hostable, Claude-Code-style harness whose default backend is an open model - Kimi-K3, served through the Hugging Face router. The design goal is not "another AI IDE." It is data sovereignty. Your source, your prompts, and the agent's intermediate reasoning stay on infrastructure you control.

Why on-prem is the whole point

Most coding assistants are thin clients in front of a closed API. That is acceptable right up until your source code is the asset you cannot leak - banks, hospitals, defense, public-sector, anyone under strict data-residency rules. For those teams, "upload the repository to a third-party endpoint" is a non-starter.

VICODE flips the default:

	Typical hosted agent	VICODE
Backend model	Proprietary API	Open model (Kimi-K3 default)
Where your code goes	Vendor cloud	Your own infra
Self-hostable	No	Yes
Model swappable	Rarely	Yes

Because the default backend is an open model, you can run the whole loop - edit, run, read errors, iterate - inside your network with no mandatory egress. Prefer a different open model? The backend is swappable.

One engineering note

Hosted reasoning models can burn their entire token budget on hidden chain-of-thought and return an empty completion. We hit exactly this and fixed it by controlling the reasoning budget (a low reasoning-effort setting) so the model spends tokens on actual code output. If you self-host an agent around a reasoning model, budget the thinking explicitly - it is an easy way to lose all your useful tokens.

CHITOS / MOUSE: the red-team half

Shipping code you did not fully write raises an obvious question: is it safe? CHITOS (also called MOUSE) is our AI security scanner and red-team agent - it probes applications and code for weaknesses rather than just generating them. There is a public demo at mouse.1street.ai so you can see the behavior before trusting it on anything real.

Together the two form a loop: VICODE writes and edits under your control, CHITOS pushes back on what was produced.

Honest scope

VICODE is a coding harness, not a magic autonomous engineer - it still needs review, tests, and a human in the loop, and open-model backends vary in capability. "No data egress" describes the on-prem deployment you run; if you point the backend at a hosted router endpoint instead, data goes wherever that endpoint lives - deploy accordingly. CHITOS/MOUSE is offered as a public demo (demo-stage): it is an assistant for finding issues, not a certification, a compliance sign-off, or a guarantee that code is secure. Treat both as tools that raise your baseline, not as replacements for security engineers or code review.

More: https://vidraft.net