French AI company Mistral published the open weights of Mistral Large 4 on Hugging Face on October 29, 2026, three weeks after a public preview. The model has roughly 1 trillion parameters in total, with 52 billion active per task, and is aimed at coding, agents and cybersecurity work.
Mistral Large 4 key facts
- Architecture: mixture-of-experts, about 1 trillion parameters, 52 billion active per task.
- Inputs and outputs: accepts text and images, produces text and code.
- Context window: 1 million tokens, according to VentureBeat.
- API price: $0.70 per million input tokens and $2.10 per million output tokens.
- Training: 3,800 Nvidia Grace Blackwell GPUs in Mistral’s European data centers.
- Hardware to self-host: a quantized version runs on four Nvidia B200 or B300 GPUs.
Open-weight license: what the catch is
The weights are available under a custom Mistral license based on MIT. According to VentureBeat, there is no revenue threshold, and internal use is exempt. However, a company that runs a model-as-a-service business, reselling access to the model, needs a separate commercial agreement with Mistral. For most businesses that want to run the model on their own servers, that makes it one of the more permissive options at this size.
The weights can be downloaded from Hugging Face, and the model is also offered through Mistral Studio and its API, OpenRouter, Microsoft Foundry and Nvidia-hosted endpoints. Self-hosting is supported through vLLM or SGLang.
Benchmarks: company claims versus independent tests
Mistral says the model scores 41% on Terminal-Bench 4.0, up from 28% at preview, ahead of GLM-5.3 at 40%. It also reports 19.7% on FrontierCode 2.0. Independent testing by Artificial Analysis found lower figures on those two tests, 25% and 15%, so the company’s numbers should not be taken as the whole picture.
Artificial Analysis rates Mistral Large 4 as the highest-scoring open-weight model developed outside China on its Intelligence Index, though it sits six points behind DeepSeek V4.1 Flash and GLM-5.3. On its Coding Agent Index the model scores 48, using about 8.5 million tokens per task against 15.9 million for GLM-5.3. The cost of running the index tasks fell 63% from the preview to $2.06, although cheaper rivals exist.
Why open-weight AI matters for developers and startups
Open weights let companies run a frontier-scale model on their own infrastructure and keep sensitive data in-house, which matters for banks, telecoms and public bodies in Europe, Africa and the Middle East. Developers can also fine-tune the model for local needs. Mistral says it supports more than 160 languages, and Mistral’s CEO Arthur Mensch has pointed out that the company has raised far less money than rival labs.
Smaller distilled versions, including one meant to run on a laptop, are planned for the coming months, but no dates have been given. For the wider race over AI spending, see our reports on Google’s Gemini 4 Carbon cybersecurity model and OpenAI’s reported plan to buy Google TPU chips.
Sources: The Next Web; VentureBeat; Mistral AI
Leave a Reply