Skip to content

AImagazine

AI Industry & Business

Mistral’s Trillion-Parameter Le Chonk Is Open Weight, but API Resellers Need a Paid Licence and the Benchmark Crown Is Contested

Mistral released the 1.05 trillion parameter Le Chonk as open weights, but API resellers need a paid licence and independent benchmarks tell a narrower story.

By John Paul 8 min read
Mistral Large 4 Le Chonk official release graphic from Mistral AI

Mistral released the full weights of Mistral Large 4 on 29 October 2026. The model, nicknamed Le Chonk, has 1.05 trillion total parameters, 52 billion active, a 1 million token context and native image input. The weights are on Hugging Face under a modified MIT licence with one large exception.

That exception is the story. Anyone can use, modify, fine-tune and commercialise the model, but a business that operates a model-as-a-service offering must sign a separate commercial licence with Mistral first. That covers API providers, GPU clouds selling inference and anyone who gives third parties meaningful control over inputs, parameters or training data. Internal use and embedded end-user products are exempt.

Mistral sells the release on benchmarks and sovereignty. It calls Le Chonk the best open-weight model for coding and state of the art in cyber defence. Independent numbers are more mixed. Artificial Analysis puts it at 26 on its Intelligence Index, the best open-weight model developed outside China but six points behind DeepSeek V4.1 Flash and GLM-5.3. It ranks second among open weights on the same lab’s Coding Agent Index.

Quick answer: Mistral Large 4 is genuinely open weight and cheap for its class at $0.70 per million input tokens and $2.10 per million output tokens. It is not open for API resale. Serving the weights to customers as a model service requires a paid deal with Mistral, and self-hosting needs a multi-GPU node.

What Mistral actually released

The release follows a public preview that began on 6 October. The architecture is a mixture of experts with 128 routed experts, four active per token and one shared expert. Mistral says it trained the model on 3,800 Nvidia Grace Blackwell GPUs in Europe. A 1.6 billion parameter vision encoder handles images.

Metric Preview, 6 October Open weights, 29 October
Input price $1.36 per million tokens $0.70 per million tokens
Output price $4.18 per million tokens $2.10 per million tokens
Cached input price Not listed $0.07 per million tokens
Context window 1 million tokens 1 million tokens
Artificial Analysis Intelligence Index 25 26
Output tokens per Intelligence Index task 135,000 82,000, down 39%

The price cut is large. Artificial Analysis estimates cost per Intelligence Index task fell from $5.54 to $2.06. That is close to Muse Spark Open at $2.02, but above DeepSeek V4.1 Flash at $0.60 and MiMo-V2.6-Pro at $0.26. Cheap for a European frontier model is not cheap in absolute terms.

The benchmark crown depends on the harness

Mistral’s strongest claim is first place on Cognition’s private FrontierCode 2.0 benchmark, which asks whether a maintainer would merge an agent’s change. Mistral reports 19.7%, ahead of DeepSeek V4.1 Flash at 19.3%, GLM-5.3 at 17.8% and Kimi K3 at 17.3%. That is a 0.4 point lead over DeepSeek, not a rout.

Artificial Analysis ran the same benchmark and got 15% for Mistral Large 4, against 29% for DeepSeek V4.1 Flash and 22% for GLM-5.3. Different harnesses, token budgets and reasoning settings can move agent scores by several points. A private run is not a neutral ranking.

On the Coding Agent Index, Mistral says the model ties the leading open-weight model at 48% while using 47% fewer tokens per task. Artificial Analysis confirms the 48 and the efficiency, but ranks it second behind GLM-5.3 at 49. The efficiency gain is real. The “best” label is a marketing choice.

Cyber defence is the other pillar. Mistral reports 51 on the Artificial Analysis Cyber Index and a 96.6% refusal rate on malicious cyber requests, the highest among the open models it tested. Artificial Analysis confirms the 51 but places it third among open weights, behind MiMo-V2.6-Pro at 56 and DeepSeek V4.1 Flash at 57. Mistral also keeps a stronger cyber variant behind vetting and contracts.

The licence is the real news

The Mistral Ecosystem Open-Weight License reads like MIT for most of its length. It grants permission to use, copy, modify, publish, distribute, sublicense and sell the model, and to train or modify it further. Then clause two adds the condition that matters.

“If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose,” the licence says. Model as a service means giving a third party access for inference or post-training in a way that lets them exercise meaningful control over inputs, parameters or training data.

Internal use is exempt. End-user products with model capabilities embedded inside specific features or harnesses are also exempt, as is relaying requests to models hosted by someone else. A chat app that uses the model behind the scenes is probably fine. A platform that lets customers pick the model, set parameters and send their own traffic is probably not.

This is a shift from Mistral’s earlier open releases. It follows a wider pattern in which labs keep weights downloadable while restricting the most valuable commercial use, which is reselling inference. The open-weight letter framed openness as a competitive necessity. The MEOW licence shows that open weights and open business models are separate questions.

What self-hosting a trillion parameters costs

The weights are free. The hardware is not. At FP8 precision, 1.05 trillion parameters occupy roughly 1.05 terabytes before the vision encoder, KV cache and serving overhead. Mistral’s vLLM recipe uses tensor parallelism across eight GPUs. The NVFP4 checkpoint is smaller, but still targets Blackwell-class accelerators.

For most Indian teams, the honest comparison is the API. At $0.70 input and $2.10 output per million tokens, a startup can run a serious evaluation for a few dollars. Renting an eight-GPU node to serve one model is a capital decision with utilisation risk, and a 1M context request makes the memory math worse. Self-hosting makes sense when data residency, custom fine-tuning or predictable high volume justify the fixed cost.

The price also lands close to OpenAI’s cheap tier. GPT-6.1 Terra costs $1 input and $5 output per million tokens. Mistral is cheaper on both, but it is not in the same capability bracket on independent coding and knowledge work tests.

Confirmed vs Not Confirmed

Claim Status Basis
Mistral published the weights on 29 October 2026 Confirmed Mistral blog, Hugging Face repository and X post, all dated 29 October
The model has 1.05 trillion total and 52 billion active parameters Confirmed Mistral docs and model card; the card says 49 billion active per token excluding embeddings
API price is $0.70 input, $0.07 cached input and $2.10 output per million tokens Confirmed Mistral docs pricing table
API resellers need a separate commercial licence Confirmed Mistral Ecosystem Open-Weight License, clause 2
Le Chonk is the best open-weight coding model Not confirmed Mistral’s private FrontierCode run says first; Artificial Analysis ranks it second on its Coding Agent Index and scores it 15% on its FrontierCode run, below DeepSeek and GLM
It is state of the art in cyber defence Not confirmed Mistral’s own chart puts it at 51 on the Cyber Index, behind MiMo-V2.6-Pro at 56 and DeepSeek V4.1 Flash at 57

What it means for Indian developers

Start with the use case, not the benchmark. If you are building an internal coding assistant, a document workflow or a product feature that calls the model behind the scenes, the licence is close to permissive and the API price is low enough to test properly. If you run an inference marketplace, an API aggregator or a GPU cloud that lets customers choose models and control parameters, talk to Mistral before you build the billing page.

Second, do not inherit a benchmark claim as a procurement decision. Run your own evaluation on your repository, language mix and latency budget. The gap between Mistral’s 19.7% and Artificial Analysis’s 15% on FrontierCode is a reminder that agent scores move with the harness. A model that finishes second on a public leaderboard can still be the cheapest correct option for a narrow task.

Third, treat data residency as a deployment question, not a licence question. Open weights let an Indian company run the model in an Indian data centre, which can help with contractual commitments and the DPDP Act. It does not transfer responsibility for security, logging, access control or model behaviour.

The strategic read is that Mistral has shipped a strong European open model with a deliberately limited resale market. That is a legitimate business choice. It is also a reminder that “open weights” describes the file, not the permission. The Microsoft and Mistral partnership already showed how distribution and control can diverge. Le Chonk makes the same point in a licence clause.

Frequently asked questions

Can I download and run Mistral Large 4 for free?

Yes. The weights are on Hugging Face at no charge, and the licence permits commercial use, modification and fine-tuning. You still pay for hardware and engineering time, and the model is too large for a single GPU.

Can I resell Mistral Large 4 through my own API?

Not without a separate commercial licence. The MEOW licence requires a paid agreement for any model-as-a-service business. Internal use, embedded end-user features and relaying requests to another host are carved out. A platform that gives customers control over the model does not fit those exceptions.

Is Mistral Large 4 really the best open-weight coding model?

Mistral’s private FrontierCode run puts it first by 0.4 points. Artificial Analysis ranks it second on its Coding Agent Index, one point behind GLM-5.3, and its independent FrontierCode result is 15%, well behind DeepSeek V4.1 Flash at 29%. It is competitive, not clearly ahead.

How much does the API cost?

Mistral charges $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. Artificial Analysis estimates about $2.06 per Intelligence Index task, compared with $0.60 for DeepSeek V4.1 Flash and $0.26 for MiMo-V2.6-Pro.

Sources

Mistral AI, “Mistral Large 4 (Le Chonk) is now open weight,” 29 October 2026. https://mistral.ai/news/mistral-large-4-oss/

Mistral AI documentation, “Mistral Large 4,” preview 6 October 2026, pricing accessed 31 October 2026. https://docs.mistral.ai/models/mistral-large-4-0

Hugging Face, “mistralai/Mistral-Large-4-1T-A52B” model card, 29 October 2026. https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B

Mistral AI, “Mistral Ecosystem Open-Weight License,” 2026. https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE

Hugging Face, “Welcome Le Chonk: Mistral Large 4 on Hugging Face,” 29 October 2026. https://huggingface.co/blog/mistral-large4

Artificial Analysis, benchmark thread on X, 29 October 2026. https://twitter.com/ArtificialAnlys/status/2115820547103952925

Share this article

Loading the next article…

Continue reading