Mistral opened API access to Mistral Large 4, a 1-trillion-parameter model with 49 billion active, and set an open-weights release for the end of October
The mixture-of-experts model, trained from scratch in about two months on roughly 4,000 Nvidia Grace Blackwell GPUs in Mistral's European data centres, is multimodal for text and images and is priced at $1.36 per million input tokens and $4.18 per million output tokens in preview through Mistral Studio. Mistral reports 61.7% on DeepSWE v1.1 and 59.4% on SWE-Atlas-QnA and calls it the best open-weight model from the US or Europe on aggregated benchmarks; a less restricted build goes to vetted cybersecurity firms and government agencies. Reflection AI announced its own 501-billion-parameter open-weight model, Beam, the day before, also with weights due this month.
why it mattersBy November there should be two Western trillion-class open-weight models you can run on your own hardware at a fraction of frontier API prices, so the self-hosting case for agent backends is worth re-costing now rather than next year.