France's Mistral AI has just done something no frontier AI lab had done at this scale: it has handed everyone the keys to its biggest model. The full weights of Mistral Large 4 — nicknamed "Le Chonk" by the company itself — are now free to download from Hugging Face, three weeks after the model first appeared as an API-only preview on 6 October. The release is leading AI news today, 9 November, and it matters because of what "open weight" unlocks — and because of the one condition in the fine print.
What just happened
Mistral first showed off Large 4 in early October, but you could only use it through the company's own API. On 29 October, it published the complete trained model — the weights — for anyone to download and run on their own hardware. The model is available today through Mistral Studio, via the API, on OpenRouter, or as a direct weights download you can self-host with tools like vLLM or SGLang.
The nickname is Mistral's own joke about the model's size. "Chonk" is internet slang for something big and heavy, and at roughly 1 trillion parameters, Large 4 is the biggest open-weight model a European lab has ever released. It is also, by Mistral's account, the first trillion-parameter model released with open weights by any lab.
What "open weight" actually means
To understand why this is a big deal, you need to know what a model's weights are. After months of training, a model is essentially a giant file of numbers — the weights — that encode everything it learned. When a company keeps those numbers private, you can only talk to the model through its API: you send a prompt, it sends back an answer, and you never see how it works.
Open weight means the company publishes those numbers. Anyone can download them, run the model on their own computers, fine-tune it on their own data, and inspect how it behaves. It is not quite the same as "open source": the training data and the training code are not fully public. But for practical purposes, the model itself is yours to run.
That is a sharp contrast with the biggest closed models, which live behind APIs and can change, disappear, or get more expensive at any time. If you run Large 4 yourself, nobody can take it away from you.
What's inside: a trillion parameters, 49 billion at a time
Large 4 is a mixture-of-experts (MoE) model. Instead of one giant network that works on every word, it contains 128 specialist sub-networks ("experts"), and for each word it processes, a router picks just 4 of them, plus one shared expert that always helps out.
Think of it like a hospital: there are 128 specialists on staff, but only four walk into the room for each patient. The hospital is huge, but the cost of each visit stays manageable. That is how a model with about 1.05 trillion total parameters only uses around 49 billion active parameters per token (52 billion if you count the embedding and output layers).
A few more facts about what's inside:
- Natively multimodal: it handles text and images, with a separate 1.6-billion-parameter vision encoder for pictures.
- Trained from scratch on European infrastructure, on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own data centres — a point the company stresses as a matter of European AI sovereignty.
- Improved since the preview: Mistral says three extra weeks of reinforcement learning went into the released weights, with the biggest gains in coding and defensive cybersecurity.
For a deeper look at how this kind of model works under the hood, see our explainer on large language models.
The licence catch
Here is the part that decides whether companies actually adopt it. Large 4 is released under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT licence. In plain terms:
- You can use it, modify it, fine-tune it and sell products built on it — for free.
- But if your business is selling the model itself as a hosted service — a Model-as-a-Service offering — you need a separate commercial agreement with Mistral.
That single condition is the difference between this licence and a fully open one. For a startup building an app on top of Large 4, it changes nothing. For a cloud provider that wants to rent out Large 4 by the hour, it means calling Mistral first. Expect the licence text, not the benchmark table, to decide how many legal departments sign off.
What it can do — and what the numbers mean
Mistral reports strong results, all of them the company's own claims until independent evaluations appear:
- 41% on Terminal-Bench 4.0, a test of real coding-agent work, up from 28% in the preview — and using 30% fewer output tokens to get there.
- 48% on the Artificial Analysis Coding Agent Index, tied with GLM-5.3, while using 47% fewer tokens on similar tasks.
- 50.9% on the Artificial Analysis Cyber Index, a defensive cybersecurity benchmark.
Mistral also published hosted API pricing: $0.70 per million input tokens, $0.07 for cached input, and $2.10 per million output tokens. The model is aimed squarely at coding agents, security triage, and companies with data-residency rules that need to run AI on their own hardware. It arrives weeks after Meta published its 816-billion-parameter Muse Spark Open model, and it sits in the same open-weight race as Kimi K3.1 Pro — see our explainer on Kimi K3.1 Pro.
Open weight vs closed API: what's the difference?
| Open weight (like Large 4) | Closed API model | |
|---|---|---|
| Who runs it | You, on your own hardware | The company, on theirs |
| Can you fine-tune it? | Yes, on your own data | No, or only through limited tools |
| Can it be taken away? | No — the weights are yours | Yes — the provider can change terms or shut it down |
| Cost | Hardware + electricity | Per-token API fees |
| Privacy | Your data never leaves your machines | Your prompts go to the provider's servers |
| Ease of use | Needs serious hardware — roughly 1 TB of weights at FP8 precision | Just an API key |
The trade-off is real: running a trillion-parameter model yourself is not a laptop job. But for companies and researchers, the control is the point.
Takeaway checklist
- Mistral Large 4 ("Le Chonk") is now open weight — the full 1-trillion-parameter model is downloadable from Hugging Face.
- Only ~49 billion parameters work on each word, thanks to a mixture-of-experts design with 128 experts and 4 active per token.
- Free to use, fine-tune and commercialise — except that selling it as a hosted service needs a separate deal with Mistral.
- Mistral claims leading open-weight coding results (41% on Terminal-Bench 4.0), but treat benchmarks as vendor claims until independent tests confirm them.
- The bigger story: open-weight frontier models are now a real alternative to closed APIs — and Europe just shipped its biggest one yet.
