On October 29 (local time), French AI company Mistral AI officially open-sourced the weights of its flagship model Mistral Large 4 (codename le Chonk), allowing developers and enterprises to take it directly, modify it, fine-tune it, and even use it for regular commercial purposes. It is based on a modified MIT license: personal, academic, modification, fine-tuning, and regular commercialization are all free; the only exception is if you want to be a "cloud provider selling compute" and do MaaS resale, in which case you must separately negotiate a license with Mistral. This "free weights, gatekeeping on resale" approach is exactly the hallmark that distinguishes Mistral from closed-source giants.
1T Parameters, Trained on Europe's Own Compute
This native multimodal model has 1 trillion total parameters and 49 billion active parameters (52 billion including embedding and output layers). It was trained from scratch entirely in Mistral's own data centers in Europe using 3,800 NVIDIA Grace Blackwell GPUs—compared with the initial version three weeks ago, this version has been optimized and upgraded after real-world testing by governments and enterprises.
Saving Tokens as a Differentiating Selling Point
The real bid to win over developers is the coding agent. On the Artificial Analysis Coding Agent Index, ML4 scores 48%, roughly on par with leading open-source model GLM-5.3, but consumes 47% fewer tokens per task on average, meaning it costs less to do the same work. Combined with online pricing ($0.70 per million input tokens, $0.07 for cached input, and $2.10 for output), Mistral is clearly aiming for the positioning of "same capability, lower invocation cost."
