Mistral's 1-Trillion-Parameter Coding Model Is Now Open Weight

Mistral released the weights of a 1-trillion-parameter model on October 29, and it's aimed squarely at coding. In the blind code tests Mistral cites, graders preferred it over every other open-weight model in the test. Only Claude Opus 5.5 beat it. You can download it and run it on hardware you control.

What happened

The release. Mistral Large 4 (nicknamed "Le Chonk") entered public preview earlier in October. On October 29, Mistral announced: "Today, we're releasing the weights of this improved Mistral Large 4." The weights are on Hugging Face as mistralai/Mistral-Large-4-1T-A52B.

The model. Per the model card, it's a natively multimodal Mixture-of-Experts model with about 1 trillion total parameters and about 52 billion active per token (49 billion without embeddings and output layers). It has a 1M-token context window and three reasoning modes (low, medium and high), and it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe.

The coding numbers (all reported by Mistral).

• Blind preference. "In Surge Code's blind pairwise evaluations, Mistral Large 4 was preferred over leading open-weight models including GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5 among the models evaluated." Its win rate was 64% against GLM-5.3 and Kimi K3, 56% against DeepSeek V4.1 Flash, and 22% against Claude Opus 5.5.
• Terminal-Bench 4.0. 41%, up 13 points from the preview.
• Artificial Analysis Coding Agent Index. 48%, which Mistral calls effectively tied with GLM-5.3 "while using 47% fewer tokens per task on average."
• FrontierCode 2.0 (Cognition's private benchmark). 19.7%, the top score on Mistral's chart.

Mistral's summary: it "ranks as the best open-weight model for coding."

What it costs

The weights are free to download. If you'd rather not host it, the Mistral API charges $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. Mistral also lists Mistral Studio, OpenRouter, Microsoft Foundry and NVIDIA-hosted endpoints.

The catch

It's big. The model card's vLLM serve command uses --tensor-parallel-size 8, and the SGLang section calls its matching --tp-size 8 an "example tensor-parallel size for a node of 8 GPUs; adjust to your topology." There's also an NVFP4 checkpoint for Blackwell GPUs "with a smaller memory footprint." The model card also points to community GGUF builds for llama.cpp, but a 1-trillion-parameter model needs far more memory than a laptop has. Budget for real GPU hardware or rented GPU capacity.

The license has one big condition. It's the Mistral Ecosystem Open-Weight License, "a modified version of the MIT license." You can use, modify, fine-tune and sell things built on it, as long as Mistral's copyright and permission notice goes with every copy. But if you or an affiliate "operates a Model as a Service business," you need a separate commercial license from Mistral before commercial use. Model as a Service means giving third parties access to the model for inference or post-training ("for instance via API or hosted endpoints") in a way that lets them meaningfully control the inputs, parameters or training data. End-user products that only embed the model inside specific features are excluded. Internal use, meaning use that doesn't make the model available to third parties, is exempt too.

So reselling it as your own API needs a deal with Mistral. Running it inside a company for that company's own people doesn't.

Try it in 15 minutes

You don't need eight GPUs to see whether it's any good for your work.

1. Make an account on the Mistral API or OpenRouter and get a key.
2. Pick a real bug or small feature in a repo you own.
3. Give it the same task you'd give your current coding model, with the same context, and compare the diffs.
4. Note the tokens used, and count reasoning tokens as output. At $0.70 in and $2.10 out per million, you can work out what a month of that task would cost.
5. If it holds up, read the model card's vLLM section. That's the setup you'd run on a client's own hardware.

Here's how you get paid

My opinion, no income guaranteed: some companies can't or won't send their source code to a cloud AI provider. Think banks, defense contractors, and anyone with strict data-residency rules. Mistral's own announcement quotes Booz Allen on "client environments with strict data-residency requirements," and Belfius, a Belgian bank, on deploying it on its own infrastructure.

Those teams still want a coding agent. Sell them a private one:

• Pilot. Use the Mistral API, or GPUs rented in the client's own cloud account if their policy allows it, on a sample of their non-sensitive code, and show them the before and after on real tickets.
• Install. Deploy Mistral Large 4 with vLLM on their hardware, wire it into the coding tools their developers already use, lock down who can reach it, and keep Mistral's license notice with the install.
• Care plan. Updates, monitoring and tuning every month.

Keep the license line in mind: you're setting it up for the client's internal use. Don't host it for third parties (via API or hosted endpoints) without a commercial license from Mistral.