Mistral Large 4 is Mistral AI's new flagship model, nicknamed "le Chonk." It went into public preview on the Mistral API on October 6, 2026, and on October 29, 2026 Mistral released the open weights on Hugging Face, two days before the October 31 date it had pointed to. The released build had about three more weeks of reinforcement learning than the preview, and the API price dropped the same day to $0.70 per million input tokens and $2.10 per million output tokens.
This guide covers what Mistral has confirmed: the model's size, the license, how to get it, the current price, the benchmark results Mistral reports, and what is still unknown. Numbers come from Mistral's preview launch post, its October 29 open weights post, the Large 4 model card and the Hugging Face model page. Benchmark figures are Mistral's own claims and have not been independently checked. The page reflects the situation on October 30, 2026.
What's new (October 30 update)
- The weights are out. Mistral's October 29 post says "We release the weights of Le Chonk." The Hugging Face repository now holds the checkpoint in 54 safetensors files, tagged for vLLM, SGLang, transformers and llama.cpp. Mistral points self hosters to vLLM or SGLang.
- The license is named. It is the Mistral Ecosystem Open Weight license, which Mistral describes as a modified MIT license. Details are in the license section below.
- The price fell. The model card now lists $0.70 input, $0.07 cached input and $2.10 output per million tokens, down from $1.36, $0.14 and $4.18.
- Three reasoning modes. Low (which replaces the old None mode), Medium, and High, the default.
- New benchmark claims. Mistral reports 41% on Terminal-Bench 4.0 (13 points above the preview) and 97.1% attack resistance on Lakera's B3 benchmark (93.3% in preview).
- More places to run it. Besides Mistral Studio, the Mistral API and OpenRouter, Mistral lists Microsoft Foundry, NVIDIA hosted endpoints and an NVIDIA NIM container.
What Mistral Large 4 is
Large 4 is a mixture of experts model. Mistral's model card lists 1.05 trillion total parameters and 52 billion active per token, plus a separate 1.6 billion parameter vision encoder. The Hugging Face page explains the other figure you will see quoted: 49 billion parameters are active per token, or 52 billion counting the embedding and output layers. The context window is 1 million tokens.
It is natively multimodal (images as well as text in) and works as a hybrid "instruct and reasoning" model, so one model handles chat, step by step reasoning and agentic tool use. Mistral calls it "our largest and most capable model to date." Other confirmed details:
- Training hardware: trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in data centers Mistral operates in Europe.
- Languages: training data covers more than 160 languages, including every official EU language.
- Version tag: version 26.10. Its predecessor, Mistral Large 3, was released in December 2025.
- API features: the model card lists structured outputs, function calling, document Q&A, batching and Mistral's agents and built in tools endpoints.
- Improved since preview: Mistral says three more weeks of reinforcement learning went into the build it released as weights, so test results from the preview may not match the current model.
Release date: preview first, weights on October 29
There are two release dates to keep apart:
- Preview API: October 6, 2026. Anyone with a Mistral account can call
mistral-large-4through Mistral Studio, and it is also listed on OpenRouter. - Open weights: October 29, 2026. During the preview Mistral red teamed the model with cybersecurity leaders, vetted partners and state authorities, who got a version with "reduced moderation and expanded cyber capabilities." The public weights are the standard build. A specialized cyber variant stays limited to vetted enterprise partners such as Orange Cyberdefense, under contract. Mistral says a full technical report is coming soon.
Google's Gemini 4 Argon also went first to vetted cyber defenders, and Google did the same with Gemini 4 Carbon, announced October 28 and opened to paying users on November 10. The difference is the destination: Large 4 is now downloadable, while Argon, Carbon and GPT-6 Astra stay behind an API.
What license does it use?
The Mistral Ecosystem Open Weight license, a modified MIT license. Mistral says the model is free to use, modify, fine tune and commercialize, with one condition: a business that operates a Model as a Service offering must get a separate commercial license from Mistral before using the model commercially. That is stricter than Mistral Large 3, which shipped under Apache 2.0, so read the license file before you build a hosted product on it. Moonshot took a similar path with Kimi K3.1 Pro: its weights arrived on November 4, 2026, under a license that also asks large Model as a Service businesses to sign a separate agreement.
Mistral Large 4 pricing
These are the current API prices on Mistral's model card, cut on October 29 when the weights shipped.
| Usage | Price per million tokens |
|---|---|
| Input | $0.70 |
| Cached input | $0.07 |
| Output | $2.10 |
Read aloud: seventy cents per million input tokens, seven cents for cached input, and two dollars ten per million output tokens. As a worked example, a job that sends 1 million input tokens and gets 200,000 output tokens back costs $0.70 plus $0.42, or about $1.12, roughly half the $2.20 it cost at the old list price.
The model card still labels Large 4 as a public preview, so prices could change again. With the weights out you can also self host, and your cost becomes your own hardware or cloud bill. A model this size needs multi GPU servers, not a laptop.
If you want to chat rather than call an API, Mistral's consumer assistant is Le Chat, which Mistral now runs as part of its Vibe app at chat.mistral.ai. Mistral's October 29 post says you can also try Large 4 in Vibe Code, the coding side of that app. If you are choosing between that app and Anthropic's assistant, see our Mistral vs Claude comparison.
Benchmarks Mistral reports
All figures below are reported by Mistral. Where the October 29 build has a new number, both are shown.
| Area | Benchmark | Mistral Large 4 |
|---|---|---|
| Coding | DeepSWE v1.1 | 61.7% |
| Coding | SWE-Atlas-QnA | 59.4% |
| Coding | Terminal-Bench 4 | 28.3% preview, 41% October 29 build |
| Coding | Coding Agent Index | 49.8% preview, 48% October 29 build |
| Coding | FrontierCode 2.0 (Cognition) | 19.7% October 29 build |
| Agents | AutomationBench (657 business workflows) | 59.9% |
| Knowledge work | AA-Briefcase | 1,393 Elo |
| Cybersecurity | Cybench (40 competition challenges) | 93% |
| Vision | Dense 200 grounding | 42% (vs 41% for GPT-6 Astra) |
| Cybersecurity | Artificial Analysis Cyber Index | 51% October 29 build |
| Safety | Lakera B3 attack resistance | 93.3% preview, 97.1% October 29 build |
What Mistral says these show:
- Coding: its Coding Agent Index score puts it ahead of DeepSeek V4 Pro and Qwen3.8 Max. In a blind human evaluation run with Surge AI, it ranked second of five models (3.74 out of 5), ahead of Kimi K3 and GLM-5.3 but behind Claude Opus 5 (4.22). For the October 29 build, Mistral says Surge graders preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, behind only Claude Opus 5.5, and that it matches GLM-5.3 on the Coding Agent Index while using 47% fewer tokens per task.
- Cybersecurity: a top five place on the Artificial Analysis Cyber Index, and 82% on a test that asks a model to reproduce and then patch a real vulnerability, which Mistral calls the highest of any model.
- Knowledge work: in legal and financial evaluations run by vals.ai, Mistral says Large 4 beat GPT-6 Astra.
Notice what Mistral does not claim. It never says Large 4 beats the best closed models overall. Its headline claim is that Large 4 is competitive with the strongest open models anywhere and "significantly" ahead of any open weight model built in the US or Europe.
The cybersecurity refusal debate
The most debated part of the launch is refusals. Mistral says several closed models, including Claude Opus 5.5 and GPT-6 Astra, "score near zero" on the vulnerability reproduction test because they refuse the task. Mistral argues that defenders often have to prove a flaw is real before they can fix it.
The same post claims Large 4 refuses malicious cyber prompts more often than any other open model, measured on refusal sets drawn from JailbreakBench, StrongREJECT and AgentHarm. Both can be true. But open weights change the safety picture: once weights are public, anyone can fine tune the safeguards away. Our guide to the Preparedness Framework vs the Responsible Scaling Policy explains how OpenAI and Anthropic decide what is too risky to release. On prompt injection, Mistral cites Lakera's public B3 benchmark, where Large 4 resisted 93.3% of attacks in preview and 97.1% after the extra training. Mistral also says it saw potential for uplift to real attacks in its live preview traffic and is changing its guardrails in response.
Mistral Large 4 vs Opus 5.5, GPT-6 and Gemini
"Mistral Large 4 vs Opus 5.5" and "vs GPT 6" are two of the most common searches about the model. No independent, like for like comparison has been published by any of the labs, so here is what the official material supports:
| Question | Mistral Large 4 | Closed flagships (GPT-6 Astra, Claude Opus 5.5, Gemini 4) |
|---|---|---|
| Can you download it? | Yes, since October 29 | No, API and apps only |
| Where it claims a lead | Visual grounding, some legal and finance tasks, vulnerability reproduction | Mistral makes no claim to beat them overall |
| Coding (Mistral's own test) | Second of five, behind Claude Opus 5 | Opus 5 ranked first |
| Self hosting and EU data control | Yes | No |
Pros of Large 4 today: downloadable weights under a mostly permissive license, a 1M token context, a lower API price and a European provider. Cons: hosted resale needs a commercial license, no independent benchmarks yet, and self hosting needs serious hardware.
For the closed flagships, see GPT-6 Astra vs Claude Fable 5.1, Claude Fable vs Opus vs Mythos and our ChatGPT vs Claude vs Gemini vs Grok comparison. For everyday chat, ChatGPT, Claude and Gemini remain the easier picks.
Who should try it now
- Developers comparing models: test the API on your own prompts first; it is cheaper now, and it tells you whether downloading 1 trillion parameters is worth it.
- Security teams: the clearest target audience, but the public weights are the standard build, not the specialized cyber variant given to vetted partners.
- European organizations with data sovereignty needs: Mistral says it will run a European deployment end to end, under European law.
- Everyday chat users: no need to switch. Wait until Large 4 reaches Mistral's chat app and independent reviews appear.
What to watch next
- The full technical report, which Mistral says is coming soon.
- Independent benchmark runs on the released weights.
- Guardrail updates, which Mistral says will continue in the coming months.
- Specialized models built on Large 4, which Mistral says are coming.
We'll update this guide when independent results appear.
From Ahzy Capital (our own products, labeled as such): Get the free AI Agent Cheat Sheet: 24 copy and paste AI agent setups for 8 types of small business, with prompts and workflows. We make and sell these, so they aren't independent picks.
Another alternative to the big US labs is AI21 Labs, and our guide to what AI21 Labs does explains its custom model business.
FAQ
When is the Mistral Large 4 release date?
The preview API launched on October 6, 2026. The open weights followed on October 29, 2026, on Hugging Face, two days ahead of the October 31 date Mistral had pointed to.
How much does Mistral Large 4 cost?
$0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens since October 29, down from $1.36, $0.14 and $4.18. You can also self host the weights and pay only for your own compute.
Is Mistral Large 4 on OpenRouter?
Yes. OpenRouter lists it as mistralai/mistral-large-4-0, at the same prices as Mistral's own API and with a context window of about 1 million tokens.
How big is Mistral Large 4?
1.05 trillion total parameters and 52 billion active (49 billion excluding embedding and output layers), plus a 1.6 billion parameter vision encoder and a 1M token context window.
Is Mistral Large 4 open source?
It is open weight, not open source in the strict sense. The weights are public under the Mistral Ecosystem Open Weight license, a modified MIT license that allows commercial use but requires a separate license for businesses offering the model as a hosted service. Large 3 used Apache 2.0.
Is Mistral Large 4 better than GPT-6 or Opus 5.5?
Not overall, based on Mistral's own claims. It reports narrow wins over GPT-6 Astra on visual grounding and some legal and finance tasks, and in Surge's blind code comparison for the October 29 build it ranked behind Claude Opus 5.5. No independent head to head exists yet.


