On October 29, 2026, the French AI lab Mistral did something none of its frontier rivals will do: it gave away the weights of its biggest model. Anyone can download Mistral Large 4 — nicknamed “Le Chonk” — from Hugging Face and run it on their own servers. No API key, no vendor deciding whether your account stays open, no per-token bill.

That single fact is why this release matters more than the usual model launch. Most capable AI systems live behind an API, where the builder controls access, pricing, and the rules. Mistral just handed a trillion-parameter model to the public — with one important catch in the licence, and safety questions the company itself openly acknowledges.

What Mistral actually released

Mistral Large 4 first appeared as an API-only public preview on October 6, 2026. The October 29 release is the weights of an improved version, after three more weeks of reinforcement learning driven by preview feedback. In Mistral’s own announcement, the pitch is blunt: “We release the weights of Le Chonk, advancing the frontier of open models in agentic coding and cybersecurity.”

The model is a mixture-of-experts system with roughly 1 trillion total parameters, of which about 49 billion are active for each token — 52 billion counting embeddings and output layers. Each token routes through 4 of 128 expert networks plus one shared expert, which is why serving it costs a fraction of what a dense trillion-parameter model would. It accepts text and images, produces text and code, and carries a 1-million-token context window — enough to hold a large codebase or months of support tickets in one prompt. It is fluent in more than 160 languages, including every official EU language, and its 1.6-billion-parameter vision encoder handles images of varying sizes and aspect ratios.

Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres, after raising €3 billion in September 2026. The release is available several ways: download the weights from Hugging Face (listed as Mistral-Large-4-1T-A52B) and serve them with vLLM or SGLang; use the Mistral API or Mistral Studio; or go through OpenRouter, NVIDIA-hosted endpoints with an optimised NIM container, or Microsoft Foundry — which Microsoft says makes it the first cloud provider to carry the model. A quantised NVFP4 checkpoint targets Blackwell GPUs, and an embedded DSpark speculative-decoding draft head was measured at up to 1.9 times higher throughput than serving the target model alone. See Unite.AI’s report on the Hugging Face release and TestingCatalog’s summary of the coding and security claims for a full rundown.

The numbers that matter — and who reported them

Mistral’s headline claim is that Large 4 is the best open-weight model for coding. Its own reported scores are:

BenchmarkLarge 4 scoreWhat it means
FrontierCode 2.0 (Cognition)19.7%Leads listed open models on repository-level coding
Artificial Analysis Coding Agent Index48%Tied with GLM-5.3, using 47% fewer tokens per task
Terminal-Bench 4.041%Up from 28% in the preview
Artificial Analysis Cyber Index51%Finding, reproducing and patching vulnerabilities
Malicious cyber request refusal96.6%Average across three safety test sets
B3 Agent Security Benchmark97.1%Attack resistance, up from 93.3% in preview
BrowseComp84.5%Web research, up from 75% at an earlier checkpoint
Finch (finance workflows)73.3%Second among evaluated open models, behind Kimi K3
Artificial Analysis Intelligence Index v526Highest among Western open-weight models

Two qualifications sit behind that table. First, nearly every figure above is Mistral’s own measurement, not an independent reproduction — the full technical report is still, in Mistral’s words, “coming soon.” Second, the company added three reasoning modes — low, medium and high, with high as the default — so developers can trade performance against cost.

On pricing, the hosted route is now $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens — a steep drop from the preview’s $1.36 and $4.18. For a business running coding agents at volume, the token-efficiency claim — 48% on the Coding Agent Index using 47% fewer tokens per task than GLM-5.3 — may matter more than any single benchmark win, because it shows up in the monthly bill.

What “open” really means here

Large 4 is released under the Mistral Ecosystem Open-Weight licence, a modified MIT licence. In plain terms: you may use, modify, fine-tune, and build commercial products on it, free of charge. The one carve-out is that businesses operating a model-as-a-service offering — selling access to the model itself as a hosted service — must obtain a separate commercial licence from Mistral first.

That distinction is fuzzier than it looks. As AI Weekly notes, the announcement does not define where wrapping the model into a software product ends and running a model-as-a-service business begins. If you plan to offer Large 4 to customers as a hosted feature, get the licence question answered in writing first. If you run it inside your own company, the open terms are straightforward.

This is also what “open weights” does not mean. It does not mean open training data or open training code, and it does not mean the model will stay unchanged. Once weights are public, Mistral cannot take them back — a point its vice president of science, Pierre Stock, made to Reuters, as reported by PYMNTS, saying he was “not comfortable living in a world in which open weights do not exist,” because once released, “it’s impossible to contain access.” That irreversibility cuts both ways: users cannot be cut off, but safety problems cannot be patched by revoking access.

Why a business should care

For most companies, the practical value comes down to three things.

Control over your data. A model on your own infrastructure keeps prompts, code, and documents inside your boundary — a real advantage for regulated or confidential work. Weights you hold also cannot be repriced or withdrawn, which protects you from API price changes, rate limits, and policy shifts. WebReOn’s own WebReOn Tools run entirely in the browser for exactly this reason: your work stays with you.

Customisation. Open weights can be fine-tuned on your codebase, documentation, and house style in ways an API model cannot — a support team on its resolved tickets, a development team on its repositories.

Cost at scale. The API price is competitive, but at high volume, self-hosting on hardware you already own can beat per-token pricing.

None of this means every business should self-host a trillion-parameter model tomorrow. The hardware bar is serious: even with only 49 billion active parameters, this is a large model that needs substantial GPU infrastructure. For most small and medium businesses, the sensible path is the hosted API, Microsoft Foundry, or an NVIDIA endpoint first — self-hosting only when data control or volume justifies the effort. Teams still judging what they need can start with free tools; our guide to the best free AI tools for students is a useful baseline.

Where the claims get complicated

The independent picture is more mixed. The Decoder’s analysis of Artificial Analysis data finds that Large 4’s 51% on the Cyber Index ranks third among open models — behind DeepSeek V4.1 Flash at 57% and MiMo-V2.6-Pro at 56% — while Kimi K3.1 Pro leads the whole index at 64%. On the broader Intelligence Index v5, Large 4 scores 26: the best Western open-weight result, but behind the leading Chinese open models overall. And closed models still lead outright: Claude Opus 5.5 scores 46% on FrontierCode 2.0 against Large 4’s 19.7%, and tops the Coding Agent Index at 62 points.

In blind human-preference comparisons run by Surge Code and cited by Mistral, Large 4’s code was preferred over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash — but it won only 22% of matchups against Claude Opus 5.5.

There is also a timing risk. Kimi’s K3.1 Pro — the model we covered when it launched on October 25 — is due to release its own open weights by November 4, and The Decoder reports it tops Terminal-Bench 4.0 at 67%, far above Large 4’s 41%. If that release lands as promised, Large 4’s open-weight coding crown could last about a week. Treat this as a fast-moving leaderboard, not a settled verdict.

The safety question Mistral is not hiding

The most interesting part of this release may be what Mistral admitted. The company says it monitored cybersecurity-related use in live traffic after the preview: the model refused black-box attacks against live targets, but Mistral also sees “potential for uplift to conduct real attacks” and is modifying its guardrails, with further updates promised.

That frames the dual-use problem clearly. A model good at finding, reproducing, and patching vulnerabilities is, by construction, also good at finding vulnerabilities. Mistral’s answer is a split: the public model keeps strong defensive capability while refusing an average of 96.6% of malicious cyber requests in its own tests, while a specialised variant built for penetration testing and red-teaming goes only to vetted partners such as Orange Cyberdefense, under contract, restricted to testing owned or explicitly authorised systems. Skimwire’s breakdown notes the same restriction.

For business readers, the practical reading is this: Large 4 is a defensive tool, not an offensive one, and any vendor claiming otherwise is misrepresenting the release. Security teams can use it to find and patch their own flaws faster — the same use case as Google’s restricted Gemini 4 Carbon model, covered earlier this week, except Large 4’s weights are public, not gated. Companies should run their own safety evaluation before putting any open-weight model near production systems, and never use it to probe systems they do not own or have explicit permission to test.

What to do now: a practical checklist

  • Try before you host. Test Large 4 via the Mistral API, Mistral Studio, or Microsoft Foundry on your own tasks, and compare it with what you use today on your own benchmarks.
  • Run the efficiency test. The 47%-fewer-tokens claim is where the money is. Measure tokens per completed task on your real workload, in all three reasoning modes, before assuming savings.
  • Check the licence. Internal use and product-building are free; reselling hosted access to the model itself needs a separate commercial licence. Get clarity before building a service on top of it.
  • Do not self-host on hope. Price the GPU infrastructure, the checkpoint, and the engineering time. For most teams, hosted access wins until volume or data rules say otherwise.
  • Wait a week before crowning a winner. Kimi K3.1 Pro’s open weights are due November 4. If you are choosing an open-weight coding model, compare both before committing.
  • Keep security evaluation in-house. Test refusal behaviour and patch quality on your own repositories, and treat vendor safety figures as claims to verify.

Frequently asked questions

Can I really download Mistral Large 4 for free?

Yes. The weights are published on Hugging Face under the Mistral Ecosystem Open-Weight licence, free to use, modify, fine-tune, and commercialise, with the single exception that model-as-a-service businesses need a separate commercial licence. You still need serious hardware or a cloud instance to run it.

Is it the best AI coding model available?

Not outright. Mistral claims it is the best open-weight model for coding, and it does lead open models on FrontierCode 2.0 — but closed models such as Claude Opus 5.5 still score substantially higher, and Chinese open models lead on several aggregate measures. Its strongest argument is coding quality per token spent, not raw capability.

Should my business self-host it?

Probably not yet, unless you have a data-control requirement or very high volume. Start hosted, measure real performance and cost, and self-host only when the numbers justify it.

Is it safe to use for cybersecurity work?

For defensive work on your own systems — finding, reproducing, and patching vulnerabilities — that is what it is built for. It must not be used to probe systems you do not own or lack explicit permission to test, and the specialised penetration-testing variant is restricted to vetted partners.

The bottom line

Mistral Large 4 is the most significant open-weight release of the year so far. A European lab has published a trillion-parameter model that anyone can download, fine-tune, and run — with coding and defensive-security capabilities that are genuinely competitive, token efficiency that could cut real costs, and a licence that is free for almost everyone except direct model resellers.

But “open” is not “best,” and “public” is not “proven.” The independent numbers complicate the headline, the safety trade-offs are real and acknowledged by Mistral itself, and a rival’s weights may reset the leaderboard within days. The smart move is the boring one: try it on your own work, measure cost per finished task, check the licence, and only then decide whether it belongs on your servers or in your toolkit. If you want help with that kind of build-versus-buy decision, WebReOn Chat and the WebReOn team work with businesses on exactly these choices every day.

Either way, the direction is clear: the frontier is no longer only something you rent by the token. Increasingly, it is something you can download — and that changes who gets to build with the best tools.