Mistral Large 4 weights are open: 1T parameters, $0.70/$2.10 on the API

Mistral released the weights of Mistral Large 4, which it also calls Le Chonk, on 29 October 2026. The model went into public preview on 6 October as an API-only product with weights promised by the end of the month. The company's post says the released version has had three more weeks of reinforcement learning since the preview, and that the same version is what the Mistral API now serves as mistral-large-4.

What was released

• Size. The Hugging Face model card describes a mixture-of-experts model with about 1 trillion total parameters and about 52 billion active per token (49 billion before embeddings and output layers), using 128 routed experts with 4 active at a time.
• Inputs and context. Text and images in, text out, with a 1 million token context window, function calling and strict JSON schema enforcement, per the card.
• Reasoning modes. Three settings per request: low, medium and high. High is the default. Low replaces the preview's "none" mode.
• Serving. Mistral recommends vLLM and also lists SGLang, Transformers and llama.cpp. A quantised NVFP4 checkpoint targets Nvidia Blackwell GPUs.
• Hosted options. Mistral Studio and the Mistral API, OpenRouter, Nvidia-hosted endpoints, and Microsoft Foundry, which Microsoft describes in the post as the first cloud provider to carry it.

The licence has one condition

The weights ship under the Mistral Ecosystem Open-Weight licence, which the licence text calls a modified version of MIT. Use, modification, fine-tuning and commercial use are free. The condition: if you or an affiliate operate a "Model as a Service" business, you need a separate commercial licence from Mistral before any commercial use.

The licence defines that as giving a third party access to the model for inference or post-training, for instance through an API or hosted endpoint, in a way that gives them meaningful control over inputs, parameters or training data. It excludes internal use, end-user products where the model sits inside specific features, and relaying requests to models hosted by others. A company that sells a product built on the model appears to be outside the condition; a company that resells inference is inside it. Have counsel read the affiliate clause if any part of your group sells hosted models.

Price and Mistral's numbers

Mistral prices the API at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. The benchmark claims are all the vendor's selection:

• Artificial Analysis Coding Agent Index: 48%, which Mistral says ties the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task.
• Terminal-Bench 4.0: 41%, up 13 percentage points on the preview, with 30% fewer output tokens.
• Cognition's private FrontierCode 2.0: 19.7%, which Mistral says leads open-weight models.
• Artificial Analysis Cyber Index: 51%, with a 96.6% average refusal rate on malicious cyber prompts from three public test sets.
• B3 agent security benchmark: 97.1% of attacks resisted, against 93.3% for the preview.

The model card is more candid than the post on one point. It reports an Artificial Analysis Intelligence Index v5 score of 26, which it calls the highest among western open-weight models, and says the leading open-weight models from China score higher on that aggregate.

What is missing

• Hardware. The card's vLLM example splits the model across 8 GPUs but states no memory requirement, throughput or cost per token for self-hosting.
• Independent results. Mistral cites third-party benchmarks, but the choice of which to show is its own, and there is no comparison with closed models beyond one human-preference test in which it says only Claude Opus 5.5 ranked higher.
• Misuse risk. Mistral says it monitored cyber-related traffic during the preview, saw "a potential for uplift to conduct real attacks", and is modifying its guardrails. The post does not say how those guardrails apply to copies of the weights that other people host.

For a team with data-residency constraints, the practical change is that a model Mistral trained in Europe can now run inside your own perimeter under a licence that permits it. For everyone else, $0.70 and $2.10 is the number to put in the routing spreadsheet, next to your own evals.