Images in this article are AI-generated concept illustrations, not screenshots from Mistral's products.
France's Mistral AI has just done what the biggest AI labs usually refuse to do: it has released the full weights of its flagship model, Mistral Large 4, to the public. Announced on October 29, 2026, the release — nicknamed "le Chonk" for its sheer size — lets anyone download a 1-trillion-parameter AI and run it on their own machines, instead of renting access through someone else's servers. Mistral pitches it as the best open-weight model yet for coding and defensive cybersecurity. Here is what Large 4 is, what changed since its October preview, and how you can actually try it.
At a glance: Open weights released Oct 29, 2026 · 1 trillion total parameters, 52 billion active per token · Natively multimodal, 1-million-token context, 160+ languages · Improved over the Oct 6 preview after 3 extra weeks of reinforcement learning · Free to download under a modified MIT-style licence · API from $0.70 per million input tokens · Runs self-hosted on as few as 4 NVIDIA B200/B300 GPUs (quantized version) · Available via Hugging Face, Mistral Studio, OpenRouter, NVIDIA NIM and Microsoft Foundry.
What Is Mistral Large 4?
Mistral Large 4 (ML4) is Mistral AI's largest and most capable model to date — a natively multimodal mixture-of-experts system. "Mixture of experts" means the model holds about 1 trillion parameters in total but only activates around 52 billion of them for any given token, so it can carry the knowledge of a giant model while costing far less to run than its size suggests. It handles text, images and code, works in more than 160 languages, and can take in up to 1 million tokens of context in a single pass — enough for very large codebases or document sets.
Mistral trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own data centres in Europe, and the company frames the release as a sovereignty play: European AI that organisations can run themselves, under their own rules, rather than depending on American or Chinese providers. Reuters reports Mistral describing Large 4 as its most capable openly available model for cybersecurity use, competing in a field dominated by Chinese open models.
What Changed Since the October 6 Preview?
Large 4 is not brand new — Mistral first showed it as a public preview on October 6, 2026, with the weights promised by the end of the month. What landed on October 29 is an improved version of that model, refined with three more weeks of reinforcement learning and feedback from governments and corporate users who tested it. The headline changes:
| What | Oct 6 preview | Oct 29 open release |
|---|---|---|
| Weights | API only; weights promised | Downloadable from Hugging Face |
| Terminal-Bench 4.0 (agentic coding) | 28% | 41% |
| API price (input / output, per 1M tokens) | $1.36 / $4.18 | $0.70 / $2.10 (about half) |
| Licence | Unnamed at preview | Mistral Ecosystem Open-Weight Licence (modified MIT) |
| Where to get it | Mistral Studio API | Hugging Face, Mistral Studio, OpenRouter, NVIDIA NIM, Microsoft Foundry |
Why Open Weights Matter — And Why Mistral Cares
Most frontier AI is closed: you send your data to a provider's servers and hope the service stays up, stays affordable and keeps your data private. Open weights flip that. Once the model files are public, anyone can download them, run them behind their own firewall, fine-tune them and keep using them even if the vendor changes its terms. That is why security teams care so much — they can point a powerful model at their own code and vulnerabilities without sending sensitive code to a third party.
"I'm not comfortable living in a world in which open weights do not exist," Mistral's Vice President of Science, Pierre Stock, told Reuters. "Once the model is out, it's impossible to contain access." That protects users from being cut off, he added — but it also means anyone can use it.
The trade-off is real and worth stating plainly: the same openness that lets defenders scan their own systems also puts a powerful model in everyone's hands. Mistral's answer is that its safety testing found a 96.6% average refusal rate on malicious cyber requests across JailbreakBench, StrongREJECT and AgentHarm — the highest it has measured among open models — and that defensive use is the point of the release.
The Benchmarks — With One Honest Caveat
Mistral's reported scores position Large 4 at the top of the open-weight field for coding and cyber defence. Treat these as vendor claims until independent testing lands — but they are specific, and they are the numbers the whole open-model community will now try to reproduce:
| Benchmark | Large 4 score (Mistral-reported) | Why it matters |
|---|---|---|
| Artificial Analysis Coding Agent Index | Top open-weight score | Full software-engineering task completion, not just code snippets |
| Terminal-Bench 4.0 | 41% (up from 28% at preview) | Agentic coding inside a real terminal |
| FrontierCode 2.0 | 19.7%, top open-weight score | Hard, research-level coding problems |
| Artificial Analysis Cyber Index | 51% | Defensive cybersecurity capability |
| Token efficiency vs GLM-5.3 | 47% fewer tokens per task | Cheaper long agent sessions — cost compounds fast |
Reality check: Every benchmark figure above comes from Mistral itself — no independent replication has been published yet at the time of writing. Also note the rollout detail: reports on launch day noted a quantized NVFP4 checkpoint available for self-hosting while the main checkpoint page still showed a short wait, so if the full weights are not yet showing on Hugging Face when you check, the quantized version and the API routes below are the working paths. And a trillion parameters is still heavy: even the quantized version needs serious GPU hardware (Mistral cites 4 NVIDIA B200 or B300 GPUs), so "run it yourself" realistically means a company or a well-equipped lab, not a laptop.
Licence and Price — The Two Details That Decide Everything
The licence is the Mistral Ecosystem Open-Weight Licence, a modified MIT licence: free to use, modify, fine-tune and commercialise — with one condition. Companies that resell the model as a Model-as-a-Service offering need a separate commercial licence from Mistral. For everyone else — developers, researchers, companies running it internally — the weights are free.
For teams that would rather not host it, the API is now roughly half the preview price: $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens through Mistral Studio, with the same model served via OpenRouter, NVIDIA NIM containers and Microsoft Foundry — Microsoft is the first cloud provider to offer it, according to Mistral.
How to Try Mistral Large 4
- Fastest no-install route: try it in Mistral's Le Chat or Mistral Studio — the model is served as mistral-large-4 through the Mistral API.
- Download the weights: get the open weights from Hugging Face (look for the NVFP4 quantized checkpoint if you want the self-hosting-friendly version) and serve them with vLLM or SGLang.
- Via a cloud or gateway: use OpenRouter, NVIDIA NIM containers, or Microsoft Foundry if you want managed hosting without running GPUs yourself.
- Check the licence first if you plan to resell access: Model-as-a-Service use needs a separate commercial licence from Mistral.
Read the original reporting at Reuters ↗, the full breakdown at Unite.AI ↗, and Mistral's own announcement at Mistral AI ↗.
FAQ
Is Mistral Large 4 really free?
The weights are free to download, use, fine-tune and commercialise under the Mistral Ecosystem Open-Weight Licence. The catch: running a trillion-parameter model needs serious GPU hardware, and reselling it as a service needs a separate commercial licence. The hosted API is paid, from $0.70 per million input tokens.
How is this different from the October 6 announcement?
October 6 was a public preview: API access only, weights promised later. October 29 is the actual open-weights release — an improved model after three more weeks of training, with lower API pricing and the licence finally published.
Can I run it on my own computer?
Realistically, no — not on a normal PC or laptop. Even the quantized version is built for around 4 high-end NVIDIA B200/B300 GPUs. For most people, the practical route is the API, Le Chat, or a cloud gateway like OpenRouter or Microsoft Foundry.
Is it better than GPT-6 or Claude?
Mistral claims the top open-weight scores for coding and cyber defence, but the best closed models still lead on several independent measures, and none of Large 4's launch benchmarks have independent replication yet. Its real advantage is openness: you can run it yourself, behind your own firewall.
The Bottom Line
Mistral Large 4's open release is the biggest open-weight moment of 2026 so far: a trillion-parameter, European-built flagship that anyone can download, with coding and cyber-defence scores that — if independent tests confirm them — put it at the front of the open field. The honest framing is that the benchmarks are Mistral's own, the hardware bar is high, and the main checkpoint rollout is still settling. But the direction is unmistakable: the most capable AI is no longer only rentable. Increasingly, it is downloadable. For defenders, developers and anyone worried about depending on a single provider, that is the story.
Would you run a trillion-parameter model yourself — or is the API enough? Tell us in the comments below!
