Europe's trillion-parameter flagship can now be downloaded and run in-house — but independent tests are more measured than the company's claims.
October 31, 2026 — Mistral AI has published the open weights of Mistral Large 4, nicknamed "Le Chonk", turning its October 6 preview into a model organisations can download and run themselves. The weights arrived on October 29, 2026, after a planned October 27 date slipped.
Image: An original illustration of a large open-weight model as a downloadable glowing core, set against a European data-centre backdrop. Credit: AI-generated illustration / Red Eye Eagle News.
Editor's Highlight: The news is the shift in control: downloadable weights users can host and customise — while independent tests show a cheaper, more efficient model, not the world's strongest.
Table of Contents
- What Was Released
- What Changed Since the Preview
- The Licence Catch
- Benchmarks: Claims vs Independent Tests
- The Security Trade-Off
- What It Means for Businesses
- What to Watch Next
- Bottom Line
- FAQ
- Sources
Key Facts
| Detail | Information |
|---|---|
| Model | Mistral Large 4 ("Le Chonk"), by Paris-based Mistral AI |
| Weights release | October 29, 2026, after an API-only preview on October 6 |
| Size | About 1.05 trillion total parameters; 49 billion active per token |
| Context window | 1 million tokens |
| Licence | Mistral Ecosystem Open-Weight License (modified MIT) |
| API price | $0.70 per million input tokens; $2.10 per million output |
| Independent score | Artificial Analysis Intelligence Index v5: 26 at high reasoning |
What Was Released
An open-weights release means the trained weights themselves are available to download, not just accessible via Mistral's API. The weights are listed on Hugging Face as mistralai/Mistral-Large-4-1T-A52B.
It is also served through Mistral Studio, the Mistral API, OpenRouter, Microsoft Foundry and NVIDIA endpoints, and can be self-hosted with vLLM or SGLang.
What Changed Since the Preview
The October 6 preview was API-only. The October 29 release adds downloadable weights, plus what Mistral describes as roughly three extra weeks of reinforcement learning and customer and security-partner feedback.
Price has roughly halved: API input now costs $0.70 per million tokens and output $2.10, against preview prices of $1.36 and $4.18. Artificial Analysis, via VentureBeat, puts the cost per Intelligence Index task at $2.06, down from $5.54 — a 63% reduction, though still above DeepSeek V4.1 Flash ($0.60).
The model is a natively multimodal mixture-of-experts design with a 1-million-token context window and three reasoning modes (Low, Medium, High), trained from scratch in Europe on 3,800 NVIDIA Grace Blackwell GPUs, according to Mistral.
The Licence Catch
Mistral Large 4 is released under the Mistral Ecosystem Open-Weight License, a modified MIT licence — not plain MIT or Apache 2.0.
Users are free to use, modify, fine-tune and commercialise the model — with one major condition. Businesses operating a Model-as-a-Service offering, giving third parties inference or post-training access, must obtain a separate commercial licence from Mistral first. There is no revenue threshold. Exemptions cover internal use and use through Mistral's official products.
Practical Takeaway: downloading weights does not make hosting free — a trillion-parameter model needs data-centre-grade hardware, with server recipes pointing to 8-GPU setups.
Benchmarks: Claims vs Independent Tests
Mistral's figures and Artificial Analysis's independent evaluation do not always agree, so both are reported here. According to Mistral, the model scores 48% on the Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task, and 41% on Terminal-Bench 4.0, up 13 points from the preview. The company also claims 19.7% on FrontierCode 2.0, which it says leads open-weight models, and 51% on the Cyber Index.
Artificial Analysis's October 29 assessment is more restrained: its Intelligence Index v5 score is 26 at high reasoning, up from 25 for the preview — the highest-scoring open-weight model developed outside China, but six points behind DeepSeek V4.1 Flash 1017 and GLM-5.3 (both 32). On coding, it ranks second among open-weight entries at 48, one point behind GLM-5.3 (49), but using 8.5 million tokens per task against GLM's 15.9 million.
Elsewhere the gaps widen: Terminal-Bench 4.0 is 25% in Artificial Analysis's evaluation, not Mistral's 41%, and FrontierCode 2.0 is 15%, behind DeepSeek (29%) and GLM-5.3 (22%).
| Benchmark | Mistral's claim | Artificial Analysis |
|---|---|---|
| Coding Agent Index | 48%, tied with GLM-5.3 | 48, second to GLM-5.3 (49) |
| Terminal-Bench 4.0 | 41% | 25% |
| FrontierCode 2.0 | 19.7%, leads open-weight | 15%, behind DeepSeek, GLM-5.3 |
| Cyber Index | 51% | 51, third among open-weight |
| Intelligence Index v5 | Not the headline claim | 26, top open-weight outside China |
Analysis Note: Neither Terminal-Bench figure is universal, and the like-for-like Intelligence Index comparison is 25 to 26.
The Security Trade-Off
Mistral pitches Large 4 to cyber defenders, arguing it helps them reproduce vulnerabilities and patch them. According to the company, it refuses 96.6% of malicious cyber requests across JailbreakBench, StrongREJECT and AgentHarm prompts.
But Mistral itself concedes the risk: it monitored cyber-related use during the preview and "we see a potential for uplift to conduct real attacks." Once weights are published, they cannot be recalled — and, as VentureBeat notes, giving operators more control also reduces the provider's control later.
Mistral chief revenue officer Marjorie Janiewicz argues customers are safer with access: "We believe that the best way for the customers to protect themselves is to have access to those models and be able to establish and customize some of those guardrails themselves."
One safeguard is deliberately narrow: a specialised pen-testing variant goes only to vetted partners such as Orange Cyberdefense, under contract, for owned or authorised systems. That restriction is a company claim, not independently verified, and the variant is not publicly available.
What It Means for Businesses
For European organisations, this is the sovereign-AI argument in practical form: a flagship trained on European infrastructure, now hostable locally. Endorsements cited by Mistral describe interest and testing, not measured results.
What to Watch Next
- The full technical report.
- Independent guardrail testing.
Bottom Line
Mistral Large 4's open-weights release is a significant step for open-weight AI in Europe: a trillion-parameter flagship now downloadable and customisable, at roughly half the preview's API price. But the independent picture is more modest — second on coding, not first on FrontierCode, and behind the leading open-weight models globally.
FAQ
What does "open-weight" mean here?
The trained weights are published and downloadable, and can be used, modified and commercialised under the Mistral Ecosystem Open-Weight License. It is not "open source."
Can I run Mistral Large 4 on my home computer?
No, not the full model. It needs data-centre-grade hardware, with server recipes pointing to 8-GPU setups.
Is it the best open-weight model available?
Not overall. Artificial Analysis ranks it top outside China, but six points behind DeepSeek V4.1 Flash 1017 and GLM-5.3, and second on coding.
Is the pen-testing version publicly available?
No. It goes only to vetted partners such as Orange Cyberdefense, under contract, for owned or authorised systems.
Sources
- Mistral AI — "Mistral Large 4 (Le Chonk) is now open weight" (October 29, 2026): official announcement.
- VentureBeat — Mistral Large 4 "Le Chonk" report: independent reporting and Artificial Analysis data.
- Skimwire — Mistral Large 4 open weights report (October 29, 2026): secondary reporting.
- YouTube (TEKTHRILL) — "Mistral Large 4 Is Now Open" explainer: explainer video.
Join the conversation
Comments
Post a Comment