Mistral Large 4 Open Weights: Pricing, Benchmarks and the License Catch

Mistral released the weights of Mistral Large 4, also called Le Chonk, on October 29. The updated multimodal model adds stronger coding results and more efficient reasoning, alongside API pricing of $0.70 per million input tokens and $2.10 per million output tokens. Self-hosting now becomes a real option.

That is the important change from the October 6 public preview. Developers can download a checkpoint rather than depend entirely on a hosted endpoint. The release also comes with a custom license, so downloading the weights and reselling access to them are different decisions.

Mistral Large 4 pricing and model details

The current model documentation lists a 1-million-token context window and a mixture-of-experts architecture with 1.05 trillion total parameters, of which 52 billion are active. It also lists a 1.6-billion-parameter vision encoder.

Detail	Mistral Large 4
API input	$0.70 per million tokens
Cached input	$0.07 per million tokens
API output	$2.10 per million tokens
Context window	1 million tokens
Inputs and outputs	Text and images in, text out
Reasoning levels	Low, medium and high; high is the default
Weights	Available to download under a custom license

The docs still carry the public-preview label and the original October 6 date. The new milestone is the October 29 weight release and improved checkpoint, not a claim that every preview designation has disappeared.

What changed in coding and reasoning

In its release announcement, Mistral reports 41% on Terminal-Bench 4.0, a 13-percentage-point improvement over its preview, while reducing output tokens by 30%. It also reports 48% on the Artificial Analysis Coding Agent Index, roughly level with GLM-5.3 while using 47% fewer tokens per task on average.

Those are the figures presented by Mistral, not results from testing here. The task, harness and reasoning setting matter when comparing them with another model. They do not establish that every coding workflow becomes 30% faster or cheaper.

Still, tokens per completed task are worth watching alongside the token price. An agent can spend several rounds reading code, making changes and checking failures. Reducing unnecessary reasoning across those rounds could matter more than a small difference in the API rate.

The downloadable model card recommends high effort for complex and agentic tasks, medium for everyday work and low for quick replies. It also makes clear that the completion-token allowance includes reasoning as well as visible output. A short final answer can therefore involve substantially more generation than the reader sees.

The Mistral Large 4 license needs a careful read

The Mistral Ecosystem Open-Weight License is a modified MIT license, not unrestricted MIT. It permits use and modification, but adds a condition for businesses operating Model-as-a-Service offerings.

If the licensee or an affiliate operates such a business, a separate commercial agreement is required before commercial use of the model or derivatives, subject to the stated exceptions. Internal use and use through Mistral's official products are exempt from that requirement. The definition also excludes specific embedded end-user features and merely relaying requests to models hosted elsewhere.

That distinction is already central to the Hacker News discussion. For a product team, the useful question is whether its actual offering falls within those terms. Calling the release simply “MIT” would obscure that work.

Downloadable does not mean cheap to operate

The model card provides vLLM and SGLang serving guidance, including examples using tensor parallelism across eight GPUs. These are deployment examples, not a universal hardware minimum. It advises reducing context length or concurrency when memory is tight.

The 52-billion active-parameter figure describes computation selected by the mixture-of-experts model. It does not make the full checkpoint a 52-billion-parameter download. Hosting decisions still need to account for the complete weights, memory for concurrent requests and the context lengths users actually send.

For an enterprise with private code and an existing inference team, control over deployment may justify that work. For a small team, the API may remain the simpler starting point. The next useful comparison is the cost of a successfully reviewed change, including retries and operations, rather than the price of a million tokens alone.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn launches and complex product updates into clear stories that developers and founders want to read and share.

Book a 30-minute call.