Introducing Mistral Large 4

Summary

Mistral AI announced Mistral Large 4, a 1-trillion-parameter mixture-of-experts model with 49B active parameters. It is a single hybrid model for instruction following, reasoning, and agentic work, and takes text and image input. It is in public preview on Mistral Studio and the API now; Mistral says open weights will follow at the end of October 2026.

Key points

• 1T total / 49B active parameters (MoE), "natively multimodal".
• Trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters, on data covering 160+ languages, including all official EU languages.
• API model ID mistral-large-4-0; preview prices $1.36 input / $4.18 output per 1M tokens (as of 2026-10-06).
• Reasoning is selectable per request; Simon Willison reports two API levels, none and high.
• Mistral's own benchmark claims: 93% on Cybench, 61.7% on DeepSWE 1.1, 28.3% on Terminal-Bench 4.0, 59.9% on AutomationBench. Simon Willison reports an Artificial Analysis index of 38, just behind DeepSeek-V4.1-Flash and far above Mistral Large 3 (9).
• The announcement does not state the context window or the licence for the weights.

Why it matters

If the weights ship as promised, this is the largest open-weight model from a Western lab and Mistral's first trillion-parameter model. It is a step up from Mistral Large 3 (~675B / 41B). Licence and context length are still unknown.