Mistral Large 4 (Le Chonk): Open Weights, License, Price and Benchmarks

Mistral Large 4, nicknamed Le Chonk, is a 1-trillion-parameter mixture-of-experts model with 49 billion active parameters. Mistral AI released it as a public preview API on October 6, 2026, then shipped an improved final version and the open weights on October 29. It reads text and images, answers in text, and now costs $0.70 per million input tokens.

Mistral calls it the best open-weight model for coding and one of the most capable open models anywhere. The independent numbers back the first half of the US-and-Europe claim and complicate the global one. This guide puts Mistral's benchmarks next to the scores Artificial Analysis publishes, and covers the specs, the new price, the licence and how to run the weights. We first wrote it on launch day and updated it on November 7, 2026, after the weights landed.

The Key Takeaways

• Size: 1.05 trillion total parameters, 49 billion active per token (52 billion counting the embedding and output layers), plus a 1.6B vision encoder. Text and image in, text out, with a 1M-token context window.
• Weights: out since October 29 on Hugging Face, two days after the October 27 date Mistral gave reporters, under the Mistral Ecosystem Open-Weight license, a modified MIT licence.
• Price: $0.70 in / $2.10 out per million tokens, about half the $1.36 / $4.18 list price of the preview.
• Independent score: 26.1 on the Artificial Analysis Intelligence Index v5.0.1, the highest of any open model from the US or Europe, but behind GLM-5.3 (31.7) and Kimi K3.1 Pro (46.5).
• The real pitch: coding and cyber defense. Mistral reports 48% on AA's Coding Agent Index, level with GLM-5.3, and 51% on the AA Cyber Index, while refusing 96.6% of malicious cyber requests.

What Mistral Large 4 Is

Mistral Large 4 is Mistral AI's largest and most capable model so far, a sparse mixture-of-experts network that only switches on about 5% of its weights for each token. Mistral's own label for it is an "open-weight hybrid instruct-and-reasoning MoE with multimodal input", which means one model handles quick answers, step-by-step reasoning and agentic tool use. The final version adds three reasoning modes, low, medium and high (the default), so you can trade depth for speed and cost on each request.

Mistral announced the preview on X at 13:06 UTC on October 6. The post carries the four claims the launch hangs on: size, the open-weight ranking, the target industries and the European infrastructure.

Meet Mistral Large 4, aka Le Chonk.

• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.

Open weights release end of October.

— Mistral AI (@MistralAI) October 6, 2026

The specs below come from Mistral's docs model card and its Hugging Face model card.

Both spec disputes from launch day are now settled. The preview materials said 49 billion active parameters while the docs card switched to 52 billion within an hour, with no note. The Hugging Face card explains it: 49 billion parameters are active per token, and 52 billion once the embedding and output layers are counted. The context window gap closed too. On launch day Artificial Analysis listed 524,288 tokens and Vals 512K against Mistral's 1M; Mistral, Hugging Face and Artificial Analysis now all list 1M.

Why it is called Le Chonk

Le Chonk is internet slang for a very fat cat, and the name is a joke the community started. VentureBeat traces it to Le Chaton Fat, a fictional giant Mistral model that went viral on X and Reddit in June with fake benchmark charts. Mistral CEO Arthur Mensch replied at the time that it was "actually le gros chaton". Chief scientist Guillaume Lample told VentureBeat that Mistral enjoyed the meme and that ML4 could be read as a first version of the idea, with larger models to come.

The training run was not a joke. Mistral says it trained ML4 from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers, and serves it from the same machines. VentureBeat, CNBC and Euronews all reported 4,000 GPUs over about two months, which appears to come from the press briefing. Mistral's written announcement and its Hugging Face card both say 3,800.

Our Le Chaton Fat explainer ended with a simple test for any trending model: no weights, no API and no docs means it is probably a kitten. On launch day Le Chonk passed two of the three. Since October 29 it passes all three.

Mistral Large 4 Benchmarks: Mistral's Numbers vs Independent Ones

Mistral Large 4 outscores every US and European open model on the independent data published so far, but it is not level with the best Chinese open models. Mistral's own charts and the third-party scores tell two slightly different stories, so they are separated here.

What Mistral reports

For the final model, Mistral reports 48% on the Artificial Analysis Coding Agent Index, which it calls "effectively tied" with GLM-5.3 while using 47% fewer tokens per task. It also reports 41% on Terminal-Bench 4.0, 13 points above the preview, and a leading 19.7% on Cognition's private FrontierCode 2.0. In Surge's blind code comparisons, graders preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, and ranked it behind only Claude Opus 5.5, according to the weights announcement.

The preview numbers came first. At launch, Mistral reported 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4.0, for a combined Coding Agent Index of 49.8%, in its launch announcement. VentureBeat checked that coding chart against the live DeepSWE leaderboard, which at the time put GLM-5.3 and Kimi K3 at about 69%, and GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 around 74%.

So 61.7% was competitive, not a lead.

The other launch-day claims are all Mistral-reported. The preview scored 59.9% on AutomationBench, a set of 657 business workflows across Gmail, Google Sheets, Slack and Salesforce, and 1,393 Elo on AA-Briefcase for long-horizon knowledge work. On the Dense 200 visual grounding test it scored 42% against GPT-6 Astra's 41%, the basis for the line that it "surpasses closed frontier models on visual grounding". In a blind coding-quality test run with Surge AI, annotators rated it 3.74 out of 5, second of five models and behind only Claude Opus 5 at 4.22. For the final model, Mistral adds 84.5% on BrowseComp and 73.3% on the Finch finance benchmark, second among open models to Kimi K3.

What independent evaluators measured

Artificial Analysis gives the final Mistral Large 4 (high) a score of 26.1 on its Intelligence Index v5.0.1, a composite of nine evaluations. Version 5 uses harder tests than the v4.3.2 index that scored the preview at 38 on launch day, so the two numbers cannot be compared. On v5.0.1 it is still the top open model from the US or Europe, ahead of Meta's Muse Spark Open, and still behind the best Chinese open models.

The preview had one clear weakness on the AA page: it was very verbose, producing 200 million tokens to run the v4.3.2 index against a median of 81 million, which pushed its cost per index task to $1.13 despite a modest list price. Mistral says the final model fixes much of that, with 30% fewer output tokens on Terminal-Bench 4.0. Our guide to AI benchmarks explains what the index measures and why a single number never settles a comparison.

Vals.ai, the second independent evaluator, scored the preview at 48.05% on the Vals Index, 32nd of 44 models overall and ninth among open-weight models. Its standout result was Harvey's Legal Agent Benchmark, where the preview scored 15.83% and ranked sixth of 75, first among open models. Vals posted that ranking within minutes of the launch.

Mistral Large 4 is now the #1 open-weight model on HLAB and #9 among open-weight models on the Vals Index.

— Vals AI (@ValsAI) October 6, 2026

The honest summary: the claim that ML4 is the best open model from the US or Europe still holds on Artificial Analysis. The claim that it competes with the strongest open models globally is a stretch on general intelligence, where GLM 5.3 leads it by about six points and Kimi K3.1 by twenty. Coding is where the final model closed the gap, if Mistral's 48% on AA's Coding Agent Index holds up.

The Cyber Pitch Is the Real Differentiator

Mistral Large 4's strongest argument is cybersecurity, not general intelligence. For the final model, Mistral reports 51% on the Artificial Analysis Cyber Index, which tests whether a model can find vulnerabilities in a codebase, reproduce them and patch them without breaking anything. At the preview launch, Mistral said one of the index's tests, reproducing a real vulnerability and then patching it, gave ML4 82%, the highest of any model, and that it solved 93% of the 40 challenges in Cybench.

The interesting part is why it wins that test.

According to Mistral, Claude Opus 5.5 and GPT-6 Astra scored near zero on the same task at launch because they refuse to perform it. Mistral argues that defenders often need to prove a flaw is real before fixing it. A provider's safety filter can block exactly that work in the middle of an incident. An open-weight model running on your own servers cannot be switched off by a policy change.

How Mistral Hedges the Cyber Pitch

Mistral is also hedging that pitch. It says the final model refuses 96.6% of malicious cyber requests, the highest rate it measured among open models. It also admits a risk: monitoring of live traffic since the preview showed "a potential for uplift to conduct real attacks", and Mistral says it is tightening its guardrails in the coming months. A specialised variant for penetration testing and red-teaming is in preview only for vetted partners such as Orange Cyberdefense, under contract and restricted to systems they own or are authorized to test. Lample framed the goal on launch day as letting enterprises and governments "defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks", as quoted by CNBC.

Z.ai faced the same tension in August. It launched GLM 5.3 without weights and said on X they would be released "in stages following rigorous safety evaluations". Mistral made a firmer promise and kept it, two days after the October 27 date it gave reporters.

Is Mistral Large 4 Open Source?

Mistral Large 4 is open weight, not open source in the strict sense. Mistral published the weights on Hugging Face on October 29, along with a quantized NVFP4 checkpoint for NVIDIA Blackwell GPUs, a technical report and serving recipes for vLLM and SGLang. The weights are the improved final model, not the preview.

Lample's launch thread explained why the weights came three weeks after the preview. The reinforcement learning run behind the preview was still in progress, and Mistral planned to ship a final version alongside the weights.

Today, we are launching a preview of our new model, Mistral Large 4 (ML4), aka le Chonk 🐈.
ML4 is a 1T-parameter model with 49B active parameters, trained natively with multimodal capabilities.

It is at the frontier of open models, and by far the strongest open-weight model from the US or Europe.
The RL run behind this preview is still in flight and shows no sign of saturation -- we will release a final version before the end of the month along with the weights of the model.
🧵
1/n

— Guillaume Lample (@GuillaumeLample) October 6, 2026

The licence is the Mistral Ecosystem Open-Weight license, a modified MIT licence. Mistral says the model is free to use, modify, fine-tune and commercialise, with one condition: a business that runs a Model-as-a-Service offering must get a separate commercial licence before using it commercially. That makes it less permissive than Large 3, which ships under Apache 2.0, and Mistral's lineup is not uniform: Mistral Medium 3.5 uses a modified MIT licence of its own. Read the terms before you build a hosted product on it.

Artificial Analysis now lists Mistral Large 4 as open weights under a commercial licence. If you are comparing it with other models you can download today, the best open source AI models list covers the field.

Mistral Large 4 Pricing

Mistral Large 4 costs $0.70 per million input tokens and $2.10 per million output tokens, with cached input at $0.07. Mistral set that price on October 29 with the final model, and it is the figure on Mistral's API rate card and docs model card.

The preview was priced differently. Its list price was $1.36 input, $0.14 cached and $4.18 output, and for the first two weeks of the preview Mistral's docs card showed those struck through and halved to $0.68, $0.07 and $2.09, without saying why. The final price lands almost exactly on that early half price.

Mistral Large 4 replaces two older models. Mistral deprecated both Large 3 and Medium 3.5 on October 31, names Large 4 as the replacement for each, and retires them on April 30, 2027. Our Mistral AI pricing guide covers the Vibe subscriptions and the rest of the API lineup. Mistral benchmarks against DeepSeek more than any other lab, and our DeepSeek V4 breakdown covers that rival.

How to Try Mistral Large 4

Mistral Large 4 is available in more places than the preview was. Developers can call it through the Mistral API in Mistral Studio, on OpenRouter, through NVIDIA's hosted endpoints or in Microsoft Foundry, which Mistral says was the first cloud provider to carry it. The docs card lists structured outputs, function calling, document Q&A, batching and Mistral's built-in agent tools as supported.

For coding, Mistral points users to Vibe Code, the coding mode of Vibe, the assistant formerly called Le Chat, where you can pick Large 4 and give it a feature to build or a bug to fix. To run it yourself, download the weights from Hugging Face and serve them with vLLM or SGLang. If you mainly want to compare answers from several frontier models without managing API keys or GPUs, Fello AI puts Mistral next to Claude, GPT and Gemini in one Mac app.

Should You Switch to Mistral Large 4?

Switch if European hosting, cyber work or self-hosting is the reason you are shopping. For a team that needs its AI processed in the EU under European law, Mistral Large 4 is the strongest model Mistral has served from its own European infrastructure, and the weights let you run it on your own servers instead. For security teams tired of refusals on legitimate vulnerability work, it is the clearest open option to test. And if you are on Large 3 or Medium 3.5, you have until April 30, 2027 to move.

For everyone else, test before you commit.

On Artificial Analysis, GLM-5.3 and Kimi K3.1 Pro still score higher on general intelligence, and both are downloadable too. The Model-as-a-Service clause matters if you plan to resell access. Where Mistral Large 4 is hard to beat is price for a model this size: $0.70 per million input tokens for a 1-trillion-parameter model you can also own outright.

FAQ

The featured image for this article was created with the help of AI.