Mistral releases open weights for Large 4 model

AI · First reported 29 Oct, 14:00 · Updated 15:36

Mistral has released the weights for its one-trillion-parameter Large 4 model after three additional weeks of reinforcement learning. Independent results show strengths in coding speed and cost, while assessments differ by benchmark and place it behind some competing models overall.

Mistral released Large 4’s weights three weeks after its public preview. The multimodal model has about one trillion parameters, with 52 billion active per step.
On Terminal-Bench 4.0, the model’s score rose from 28% to 41%; Mistral says it uses 47% fewer tokens per task than GLM-5.3 on the Coding Agent Index.
Artificial Analysis rated the final model at 26 on its Intelligence Index, compared with 32 for DeepSeek V4.1 Flash and GLM-5.3. It measured Large 4’s generation speed at 161 tokens per second and average index-task cost at $2.06.
On Artificial Analysis’ Cyber Index, Large 4 scored 51%, behind two open models listed in Mistral’s comparison. Mistral reported a 96.6% refusal rate for malicious cyber requests and said it plans to strengthen safeguards.
The model is available under a modified MIT-style license; businesses offering it as a commercial cloud service need a separate license. Mistral lists API pricing at $0.70 per million input tokens and $2.10 per million output tokens.

Summary written by Lux AI from 4 articles

Key facts

Model size: About 1 trillion parameters; 52 billion active per step.
Terminal-Bench 4.0: 41%, up from 28% in preview.
Artificial Analysis Intelligence Index: 26 points for Large 4; 32 each for DeepSeek V4.1 Flash and GLM-5.3.
Cyber Index: 51%; two open models in Mistral’s comparison scored higher.
API pricing: $0.70 per million input tokens and $2.10 per million output tokens.