Mistral releases 1T-parameter Large 4 open weights, cuts API prices

Mistral released downloadable weights for its 1-trillion-parameter Mistral Large 4 model on Hugging Face, alongside an API price cut of roughly 50 percent. The model is a mixture-of-experts architecture with approximately 49 billion parameters active at any given time out of the 1 trillion total, making it more efficient to run than the parameter count suggests. The release follows three weeks of reinforcement learning added after an earlier preview, which Mistral says raised the model's Terminal-Bench 4.0 score from 28 percent to 41 percent and produced results that, on Artificial Analysis's Coding Agent Index, tie GLM-5.3 at roughly 48 percent while using 47 percent fewer tokens per task. Third-party testers ThibaultJaigu and RequestyAI assessed it as the second-best open-source model on their coding benchmark. The new API pricing is $0.70, $0.07, and $2.10 per million tokens for input, cached, and output respectively, down from $1.36, $0.14, and $4.18. The model is released under the Mistral Ecosystem Open-Weight license, which permits use, modification, and commercialization but requires Model-as-a-Service providers to obtain a separate commercial license, a tighter arrangement than the Apache 2.0 license used for Mistral Large 3. A stronger cyber variant of the model was withheld from public release entirely and given only to vetted partners such as Orange Cyberdefense for authorized penetration testing, after Mistral's own testing suggested the model could increase attackers' real-world abilities even though it refused 96.6 percent of malicious cyber requests on selected benchmarks.