Le Chonk: the 1-trillion-parameter European open model now has its weights out
Direct answer
Le Chonk is the nickname of Mistral Large 4, Mistral AI's 1-trillion-parameter model. On October 29, 2026 Mistral published its full weights on Hugging Face: any business can download the model, fine-tune it and run it on its own infrastructure. Only 49 billion parameters activate per token, thanks to its mixture-of-experts architecture. The European open route is now verifiable in hand, with public figures and a public license.
What shipped on October 29
- The full weights, downloadable: the Mistral-Large-4-1T-A52B model arrived on Hugging Face three weeks after its API-only preview opened on October 6. The download lets you run it, fine-tune it and build products on it, with the usual vLLM and SGLang tooling and an NVIDIA NIM container.
- 1 trillion parameters, 49 billion active: Large 4 is a mixture of experts with 128 routed experts, of which 4 plus 1 shared expert process each token. At 52 billion counting embedding and output layers, the model delivers frontier-class capacity at a much lower per-token compute cost than its total size.
- Multimodal with 1-million context: the model reads text and images, handles more than 160 languages, including every official language of the European Union, and works with a 1-million-token context window. Mistral trained it from scratch on its own infrastructure in Europe.
- An open license with one clear limit: Large 4 ships under the Mistral Ecosystem Open-Weight License, a modified MIT-style license. Commercial use is allowed, and selling the model itself as a hosted service requires a separate commercial agreement with Mistral.
What the numbers say (and whose they are)
- 48% on the Artificial Analysis Coding Agent Index: tied with GLM-5.3 and using 47% fewer tokens per task, according to the independent measurement Mistral itself cites. This is the strongest figure of the launch, because it comes from an outside evaluator.
- 41% on Terminal-Bench 4.0: up from 28% in the preview, with 30% fewer output tokens. The improvement between the preview and the final weights came with three more weeks of reinforcement learning.
- The rest are Mistral's claims: the company says it is the best open model on Cognition's FrontierCode coding test and that it refuses malicious cyber requests well. Those figures are the vendor's until independent evaluations land at scale.
- API pricing for comparison: through the hosted API, Large 4 costs $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens. Open weights remove the vendor, and the bill arrives in hardware and operations.
What to do in your business this week
- Price the real cost of self-hosting: the full weights run to about 1 terabyte in FP8 and around 0.5 terabytes in NVFP4, before the cache that grows with every long conversation. Get a GPU and operations quote before celebrating the zero download price.
- Use the API to test before installing: Large 4 is also available through the API in Mistral Studio, OpenRouter, NVIDIA-hosted endpoints and Microsoft Foundry. Run your real tasks through the API and measure quality and cost per finished task.
- Check the license against your business model: if your plan is to sell the model itself as a hosted service, you need Mistral's commercial agreement. If you use it inside your product, the open license covers you.
- Compare with the closed option you already use: the open model gives you auditability and data sovereignty. The closed model gives you operations without infrastructure. The decision is between control and convenience, with numbers on the table.
Note: this article summarizes Mistral Large 4's weight release of October 29, 2026, the model card on Hugging Face and coverage by The Independent AI, TechToday, CosmoXOrigin and AI News. The Artificial Analysis figures are independent measurements cited by Mistral; the remaining benchmarks are the vendor's claims. Prices are Mistral's official list prices. The November 15 topic table placed the weight release on November 13; the verified date is October 29.
Want AI that works for your business?
Message us on WhatsApp and let's talk about automation and AI agents for your business.
WhatsApp inicIARelated articles
- Mistral Large 4: the 1-trillion-parameter sovereign open model and its €3,000M round
- The open model you can run on your own server: Aleph Alpha Kolibri (Apache 2.0)
- Claude Haiku 5.5 costs up to 90% less: the AI price war continues (and it benefits you)
Frequently asked questions
What is Le Chonk?
It is the nickname of Mistral Large 4, Mistral AI's 1-trillion-parameter model. Its full weights were published on October 29, 2026 on Hugging Face, and any business can download, fine-tune and run them on its own infrastructure.
How many parameters activate per token?
Around 49 billion per token, out of 1 trillion total, thanks to its mixture-of-experts architecture with 128 routed experts. Counting embedding and output layers, the figure is 52 billion.
Can I use it commercially?
Yes, under the Mistral Ecosystem Open-Weight License, a modified MIT-style license. Selling the model itself as a hosted service requires a separate commercial agreement with Mistral.
How much does it cost to use?
Through the hosted API it costs $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens. Downloading the weights is free; the cost arrives in hardware, GPUs and operations.
