Mistral Large 4’s weights are out. What it takes to run, and how to use it now

What Mistral released

On October 29, 2026, Mistral released the weights of Mistral Large 4, the model it nicknames Le Chonk: “We release the weights of Le Chonk, advancing the frontier of open models in agentic coding and cybersecurity.” The model came out as an API-only public preview on October 6, and The Next Web reports that it had three more weeks of reinforcement learning before the weights went out, so the download isn’t the preview.

The model card on Hugging Face describes a mixture-of-experts model with about 1 trillion parameters, of which about 52 billion are active for each token (49 billion without the embedding and output layers). It reads text and images, writes text, and has a 1M-token context. It has three reasoning efforts, set with reasoning_effort: low, medium and high, the default.

The weights come under the Mistral Ecosystem Open-Weight License, which Mistral calls a modified MIT license. You may use, change, fine-tune and sell with it. The one catch: a business offering “a third party access to the Model or its derivative works for inference or post-training” must “enter into a separate commercial license with Mistral.” Internal use is exempt, as is use through Mistral’s own products. There’s no revenue threshold.

How good it is

These are Mistral’s own figures, from its announcement and as The Next Web reports them, except the last row, which is Artificial Analysis’s:

Test	Mistral Large 4	Compared with
Terminal-Bench 4.0	41% (28% at preview)	GLM-5.3, 40%
Artificial Analysis Coding Agent Index	48%	GLM-5.3, effectively tied
Artificial Analysis Cyber Index	51%
Malicious cyber requests refused	96.6%	Highest of the open models Mistral measured
Coding Agent Index, in Mistral Vibe (Artificial Analysis)	48.3 at high effort, $1.26 a task	Claude Code with Opus 5.5 leads

Mistral says it uses 47% fewer tokens per task than GLM-5.3 on the Coding Agent Index, and that in Surge’s human-preference test of coding it ranked behind only Claude Opus 5.5. On Artificial Analysis’s index, as our memo on it sets out, Mistral Vibe with Large 4 sits below the top ten, at 35.4 on medium effort. It’s a strong open model for coding, not a replacement for the best closed ones.

What it takes to run

More than any laptop or desktop has. Mistral’s serving examples on the model card run on eight GPUs, with --tensor-parallel-size 8 in vLLM and --tp-size 8 in SGLang, and there’s an NVFP4 build for Nvidia’s Blackwell GPUs. The Next Web reports that a quantized version runs on four Nvidia B200 or B300 GPUs.

For the rest of us, Unsloth’s GGUF builds work with llama.cpp, Ollama and LM Studio:

Build	Size	Unsloth’s guidance
1-bit (UD-IQ1_S)	233 GB	A device with 256 GB of memory
2-bit	302 to 378 GB	384 to 512 GB
4-bit (UD-Q4_K_XL)	605 GB
8-bit (Q8_0)	1.12 TB	1.2 TB or more

Unsloth’s Ollama command for the 4-bit build is ollama run hf.co/unsloth/Mistral-Large-4-1T-A52B-GGUF:UD-Q4_K_XL. Mistral says Large 4 “will serve as the foundation” for smaller, specialized models, and The Next Web reports that distilled versions will follow “in the coming months,” including one that runs on a laptop, without a date.

Four ways to use it in a coding agent

1. Mistral Vibe. Mistral’s own coding agent, and the setup Artificial Analysis tested. Mistral says to try Large 4 “in Vibe Code; give it a feature to build or a bug to fix in your own repository.” The CLI installs with curl -LsSf https://mistral.ai/vibe/install.sh | bash (docs). On Mistral’s pricing page, the free plan has “limited access to Vibe,” and Pro, at $14.99 a month, has “full access to Vibe for all-day coding and long-running tasks,” with $15 of API credits.
2. Mistral’s API, per token. The model is mistral-large-4-0 at $0.70 per million input tokens, $0.07 cached and $2.10 output (model card): about a third of Claude Sonnet 5.5’s or GPT-6.1 Sol’s $2 input price, and a fifth of their $10 output price. OpenCode doesn’t list Mistral as a provider yet, so add it as a custom OpenAI-compatible provider, as OpenCode’s provider guide describes. Mistral also lists it on OpenRouter, which OpenCode does list, though its page today showed only EU routing through Mistral.
3. Your own servers. Serve the weights with vLLM or SGLang from the model card, or llama.cpp with Unsloth’s builds, and point an agent at the server’s OpenAI-compatible address. OpenCode’s guide covers Ollama at http://localhost:11434/v1 and llama.cpp at http://127.0.0.1:8080/v1.
4. Hosted elsewhere. Mistral says Large 4 is also on Nvidia-hosted endpoints and Microsoft Foundry.

Enjoy runs OpenCode beside Claude Code, Codex and the other labs’ agents. Set up Mistral’s API or your own server in OpenCode once, and Mistral Large 4 appears in Enjoy’s model picker with the rest of OpenCode’s models. Hand it a conversation when a Claude or ChatGPT plan runs low, or give it the same task to compare. With a model you serve yourself, the work stays on your machines. Enjoy is a monthly subscription, with the first seven days free (pricing); Enjoy doesn’t add Mistral to a Claude or ChatGPT account.

What Mistral hasn’t said

• What it takes to run. The model card gives no memory figure; the hardware above is from Unsloth and The Next Web.
• When the smaller versions come, or how small they’ll be.
• Which Vibe plan gets how much Large 4. The pricing page doesn’t mention the model.
• The technical report, which Mistral lists as “coming soon.”

We’ll update this memo when Mistral publishes the smaller versions.