Mistral Large 4 Opens Its Weights: What Changes

Mistral released the weights of Mistral Large 4 on October 29, 2026, after offering the model through an API preview earlier in the month. Developers can now evaluate a self-hosted deployment as well as Mistral’s API. The current API listing shows $0.70 per million input tokens and $2.10 per million output tokens; the cost of running the downloaded model depends on hardware and workload, which Mistral does not price for you.

What changed since the preview

Mistral’s October 6 announcement invited developers to try a public-preview API and said the weights would arrive by the end of the month. Its October 29 announcement says the improved weights are available. That changes the deployment choice: an organization can download the model, run inference on infrastructure it controls, and assess whether that control is worth the engineering and hardware bill.

Decision point	October 6 preview	October 29 release
Access	Public-preview API; weights promised later	Downloadable weights and API access
Reasoning control	Preview model	Low, medium, and high effort on the released model; high is the default
Deployment	Mistral-hosted preview	Self-hosting becomes an option under the model license
Operating cost	Historical API price not established by the pages inspected	Current API rates are published; self-hosting cost requires your own measurement

The new model card lists low for quick responses, medium for everyday tasks, and high for complex or agentic work. Mistral says low replaces the earlier “None” reasoning mode. Changing effort can change both response quality and token use; measure those effects on your own prompts before settling on a default.

Mistral reports that the released model scored 41% on Terminal-Bench 4.0, up 13 percentage points from a rounded 28% for the preview, while using 30% fewer output tokens on that benchmark. The preview article gives the earlier score more precisely as 28.3%. These are results reported by Mistral, not tests we ran, and they do not establish the same improvement on your workload. Its reported 47% lower average token use than GLM-5.3 on the Coding Agent Index is likewise a benchmark comparison, not a promised reduction in an individual API bill.

What the API costs now

The Mistral model page currently lists a one-million-token context window and these USD rates. The inspected pages do not establish a separate maximum output length or a verified historical price cut.

Token type	Current price per million tokens
Input	$0.70
Cached input	$0.07
Output	$2.10

For an illustrative request batch with one million uncached input tokens and 200,000 output tokens, the token charge is 1 × $0.70 + 0.2 × $2.10 = $1.12, before any other applicable costs. This is arithmetic using the published rates, not a measured bill. Cached input would have a different charge only where the provider’s caching rules apply. Enter your expected token counts in the AI API cost calculator and compare the assumptions with the broader API price guide.

The one-million-token context figure is a ceiling stated in the model listing, not a reason to send every document in full. Longer prompts consume input tokens and may affect latency. The site’s context-window guide explains why useful context can be smaller than the advertised window.

What self-hosting changes

Mistral’s model documentation describes 1.05 trillion total parameters, 52 billion active parameters and a 1.6-billion-parameter vision encoder. The model card explains that its approximately 49 billion active parameters per token exclude embedding and output layers, while the approximately 52 billion figure includes them. Active parameters describe computation for a token; they do not mean the full model weights fit in memory sized for 52 billion parameters.

The card provides deployment routes using vLLM, SGLang, Transformers and llama.cpp, and points to a separate NVFP4 checkpoint aimed at Blackwell GPUs. Those are implementation paths, not measured hardware requirements for your setup. Model weight format, context length, concurrency and KV cache all affect memory demand. Use the local LLM VRAM guide to build a first estimate, then test on the exact checkpoint and serving stack you would use. The published card does not give a universal server cost or throughput figure.

The weights carry the Mistral Ecosystem Open-Weight License, a modified MIT-style license. It permits broad use and modification, but includes conditions, including a separate commercial-license requirement for a licensee or affiliate operating a Model-as-a-Service business for commercial purposes. Read the definitions and exceptions in the full license against your intended product before deployment; “open weights” alone does not settle those terms.

What to do before switching

Start with a small, representative set of your own tasks. Test correctness and the failure cases that matter to you across low, medium and high reasoning effort. Record input and output tokens, latency, and any caching actually achieved. The default high setting may be appropriate for hard tasks, while a simpler workload may justify a lower setting only after you compare results.

If data residency, customization or control of the serving stack is the main reason to download the model, test the model-card deployment recipe and measure memory use, throughput and operational effort. Compare those measured costs with an API estimate at your expected volume. If you are still deciding which model to evaluate, check the AI model tracker for other recent releases. The new weights expand your options; the best route depends on your own quality requirements and measured operating costs.

Sources

• Mistral’s open-weight announcement — release, reasoning modes and maker-reported benchmarks.
• Mistral’s preview announcement — October 6 access and earlier reported benchmark.
• Mistral Large 4 documentation — current context, architecture and API prices.
• Mistral Large 4 model card — weights and deployment guidance.
• Mistral Ecosystem Open-Weight License — exact license terms.

All pages read on October 30, 2026.

Free toolWhat will 10,000 chats a month cost?Estimate my bill