Mistral AI releases the open weights of Mistral Large 4
The weights of the model, with about 1 trillion parameters, can be downloaded from Hugging Face under a modified MIT license. On the API it costs $0.70/$2.10 per million tokens.
Source of truthmistral.ai/news/mistral-large-4-ossMistral AI has published the weights of Mistral Large 4, in an improved version nicknamed Le Chonk. The model launched in public preview earlier in October, and the company had promised the weights by the end of the month.
It is an open-weights model with about 1 trillion parameters. According to the model card, about 49 billion of them are active per token, or 52 billion counting embeddings and output layers. It takes text and images as input and has a context window of 1M tokens. It offers three reasoning modes: Low, Medium and High.
The license is the Mistral Ecosystem Open-Weight License, a modified MIT license. It allows free use, modification, fine-tuning and commercial use. The one exception: businesses that offer the model as a service (Model-as-a-Service) need a separate commercial license from Mistral AI.
The weights are on Hugging Face and run with vLLM or SGLang. Mistral AI says the model is also available in Mistral Studio, through its API, on OpenRouter, through Microsoft Foundry and on NVIDIA-hosted endpoints. On the API it costs $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens.
According to Mistral AI, the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own datacenters in Europe. The company also reports its own benchmark results, such as 41% on Terminal-Bench 4.0. It says it is changing its cyber guardrails because, since the preview, it has seen that the model could help with real attacks. A full technical report is coming soon, the company says.
For a Greek software company or an accounting firm building its own tools, one option is the API: $0.70 per million input tokens and $2.10 for output. If you want to sell it as a service, you need a commercial license. Mistral AI says the model is fluent in every official EU language, so Greek too, but gives no figures on its quality in Greek.
Other sources: mistral.ai/news/mistral-large-4huggingface.co/…/Mistral-Large-4-1T-A52Bhuggingface.co/blog/mistral-large4
Related stories
Replica: GitHub's open-source AI prototyping tool, under MIT
According to GitHub, Replica builds interactive prototypes from your team's design system, through the GitHub Copilot app.
LFM3: six small open-weight models from Liquid AI
Models in the 2B-parameter class for text, vision, audio, omni and decisions. The cards list Greek only among speech languages.