France's Mistral releases Large 4 AI model weights and cuts API prices

The model, nicknamed Le Chonk by the company, is built on a Mixture-of-Experts architecture: out of roughly 1 trillion parameters, 49 billion are activated when processing each token. It operates with a context window of up to 1 million tokens, recognizes images, and supports over 160 languages. The weights have been published on Hugging Face; additionally, the model is available via Mistral Studio, OpenRouter, and NVIDIA infrastructure. Three reasoning modes are provided — Low, Medium, and High, with the latter enabled by default.

Before the release, Mistral spent three more weeks fine-tuning the model with reinforcement learning. According to the company, its score on Terminal-Bench 4.0 rose by 13 percentage points to 41%, while the number of generated tokens decreased by 30%. On the Artificial Analysis Coding Agent Index, the model scored 48% and, Mistral claims, uses on average 47% fewer tokens per task than China's GLM-5.3. In the company's internal tests, Large 4 rejected 96.6% of malicious cyberattack prompts; however, Mistral acknowledges that the model is potentially capable of assisting in real-world attacks and pledges to further refine restrictions.

Now, 1 million input tokens via API cost $0.70, output tokens cost $2.10, and cached input tokens cost $0.07; at announcement, the baseline rates were $1.36 and $4.18. The Mistral Ecosystem Open-Weight license, based on a modified MIT license, allows free use, fine-tuning, and commercial application, but paid access under a Model-as-a-Service scheme requires a separate agreement with Mistral. According to Habr, running the FP8 version locally will require about 1 TB of disk space, while the NVFP4 build requires around 610 GB and NVIDIA Blackwell-generation graphics cards.