Mistral Large 4: Cheap AI model gets open weights with a catch

French AI company Mistral announced its new language model Large 4 as a public preview on October 6, 2026. The model has roughly 1 trillion (1,000 billion) parameters in total, but uses only about 49 billion at a time per token. This is called mixture-of-experts, and it means the model draws on only part of itself each time it responds. The sources, however, cite both 49 and 52 billion active parameters.

Pricing undercuts competitors

According to the outlet Forkast, Large 4 costs $0.68 per million input tokens and $2.09 per million output tokens. Tokens are the small chunks of text a model reads and writes, and they are used to bill for usage.

By comparison, according to Forkast, OpenAI's GPT-6.1 Sol and Google's Gemini 4 Argon both cost $2 for input and $10 for output. Microsoft's MAI-Thinking-1 costs $2 and $8. Forkast concludes that Large 4 undercuts the price floor for proprietary, meaning closed, models.

The pricing is not clear-cut, though. Let's Data Science reports that the preview version was priced at $1.36 for input and $4.18 for output. The sources do not make clear whether the difference is due to preview versus final pricing.

The model is available through Mistral's own platform La Plateforme and, according to Activepieces, also through Azure AI Studio and AWS Bedrock.

Open weights, but with conditions

Mistral promised to release the model's weights, meaning the trained model itself, which can be downloaded, by the end of October. According to Forkast, the license, the Mistral Ecosystem Open-Weight License, was published on October 29. Let's Data Science reports that the weights were posted on Hugging Face the same day.

Forkast reports that the license requires companies that operate a model-as-a-service offering to obtain a separate commercial license from Mistral before using the model commercially. The threshold depends on business model rather than revenue. Let's Data Science describes it as a modified MIT license that allows downloading and fine-tuning, while resellers of API access need Mistral's approval.

For ordinary companies that want to run or adapt the model themselves, the license thus appears to offer freedom. It is the providers that want to resell access to the model that must negotiate with Mistral.

Background

Mistral's Large 3 was released in December 2025 under the more permissive Apache 2.0 license. Medium 3.5 has a modified MIT license with an exception for high-revenue companies.

In September 2026, according to Let's Data Science, Mistral closed a Series D funding round of EUR 3 billion at a valuation of more than EUR 21 billion. Co-founder Guillaume Lample has argued for open weights as a kind of insurance: If you use a closed model, there is no guarantee it will still exist tomorrow.

According to Forkast, the benchmark results for Large 4, including 62 percent on DeepSWE and 92 percent on HumanEval, were provided by the vendor or reproduced by secondary sources.