Mistral releases Large 4 open AI model with one trillion parameters

French AI lab Mistral has released its newest model, Mistral Large 4, to the public, claiming it has caught up with Chinese rivals at the frontier of open-weight AI. The model contains one trillion parameters and became available on the open-source platform Hugging Face on Thursday.

Mistral says Large 4 narrowly tops all other open models on Artificial Analysis’ Coding Agent Index, which measures how well AI agents complete full software engineering tasks. On Terminal-Bench 4.0, a test of coding agents, the new model ties with China’s GLM-5.3, previously considered the best open model in the world.

“The company tries to compete with the American models, with the idea that you do not have to depend on the United States for AI technology.”

Benchmark figures provided by Mistral show that additional reinforcement learning training since a limited preview released on Oct. 6 lifted the model’s accuracy on Terminal Bench 4.0 from 28 percent to 41 percent. Mistral attributes the gains to three more weeks of reinforcement learning.

The company also says Large 4 is much more efficient than its Chinese open-weight rivals, using 47 percent fewer tokens per task than GLM-5.3. Mistral often calls the model “Le Chonk,” a humorous French wordplay based on its cat mascot.

Proprietary models from companies such as OpenAI and Anthropic remain ahead in overall capabilities. For example, Anthropic’s Mythos 5.1 scored 60.9 percent on the same benchmark.

Mistral has raised more than 6.6 billion euros since its founding in 2023, according to market analyst Crunchbase. By comparison, about 181 billion euros has been invested in OpenAI and 153 billion euros in Anthropic.

Experts say Mistral is the only serious European answer to American and Chinese AI companies, though its products typically arrive a few months later and are slightly less powerful.