Mistral Launches Open-Source Trillion-Parameter Large 4

Mistral AI has released Mistral Large 4, its most capable LLM to date, in public preview via its cloud platform, with model weights to follow later in October. The model uses a mixture-of-experts architecture with 1 trillion total parameters, activating only 49 billion at a time for efficiency. It supports over 160 languages, ranks in the top five on the AA Cyber Index for cybersecurity tasks, and outperforms open-source rivals including Qwen3.8 Max and DeepSeek V4 Pro. Trained on 3,800 Grace Blackwell chips, Mistral says the ongoing training run will produce larger versions in coming months.