started · updated
Mistral AI releases Mistral Large 4 open-weight model
Mistral AI has released the weights for its flagship Mistral Large 4 (ML4) model, codenamed “Le Chonk,” making it available for download on Hugging Face. The 1.05 trillion parameter multimodal model uses a mixture-of-experts architecture, activating approximately 52 billion parameters per task.
Following three weeks of additional reinforcement learning after its initial October 6 preview, the model shows significant improvements in coding and cybersecurity. It achieved a 41% score on the Terminal-Bench 4.0 benchmark, up from 28% in the preview version. On the Artificial Analysis Coding Agent Index, the model scored 48%, performing similarly to China’s GLM-5.3 while using 47% fewer tokens per task.
Designed for cybersecurity, the model can detect, analyze, and respond to threats in code and IT infrastructure. Mistral reported that the model scored 51% on the Artificial Analysis Cyber Index and successfully rejected 96.6% of malicious cyber requests during testing. The model was trained on European infrastructure using 3,800 NVIDIA Grace Blackwell GPUs.
Under a modified MIT license, the model allows for personal, academic, and commercial use, including fine-tuning. However, entities wishing to resell the model as a hosted service must obtain a separate commercial license from Mistral.
Entities
Anthropic · Cloudera · Europe · France · GLM-5 · Hugging Face · Mistral AI · Mistral Large 4 · Nvidia · OpenAI · Pierre Stock
Claims
What the coverage asserts, and how many sources carry each claim.
- [● 9 SOURCES] The model's Terminal-Bench 4.0 score rose from 28% to 41% after additional training. es.euronews.com · it.euronews.com · pt.euronews.com · novini247.com · thebridge.jp · +4 more
- [● 6 SOURCES] Mistral integrated two additional reasoning modes into the model during the reinforcement learning cycle. es.euronews.com · it.euronews.com · pt.euronews.com · dijitaliyidir.com · novini247.com · +1 more
- [● 5 SOURCES] The ML4 model is designed to detect, analyze, and respond to cyber threats in code and IT infrastructure. es.euronews.com · it.euronews.com · pt.euronews.com · novini247.com · olhardigital.com.br
- [● 23 SOURCES] Mistral released the weights for its Mistral Large 4 (ML4) model. es.euronews.com · it.euronews.com · pt.euronews.com · cryptocompass.com · mistral.ai · +18 more
- [● 15 SOURCES] ML4 uses an average of 47% fewer tokens per task than the Chinese model GLM-5. es.euronews.com · it.euronews.com · pt.euronews.com · cryptocompass.com · www.cryptopolitan.com · +10 more
- [● 20 SOURCES] Mistral Large 4 has 1.05 trillion total parameters and 52 billion active parameters. cryptocompass.com · www.smartphonology.it · www.ithome.com · olhardigital.com.br · thebridge.jp · +15 more
- [● 6 SOURCES] The model supports a context window of one million tokens. www.smartphonology.it · thebridge.jp · dijitaliyidir.com · espeo.eu · borncity.com · +1 more
- [● 5 SOURCES] The model was trained using 3,800 NVIDIA Grace Blackwell GPUs on European infrastructure. www.ithome.com · olhardigital.com.br · thebridge.jp · dijitaliyidir.com · mistral.ai
- [● 10 SOURCES] Mistral Large 4 scored 48% on the Artificial Analysis Coding Agent Index. cryptocompass.com · www.ithome.com · thebridge.jp · dijitaliyidir.com · mistral.ai · +5 more