Mistral AI released Mistral Large 4, a 1 trillion-parameter multimodal model. It uses 49 billion active parameters per token. The model was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters. It handles text and image input with text output. It supports a 1 million-token context window. It is natively fluent in over 160 languages. It offers three reasoning efforts: low, medium, and high. It includes native function calling and strict JSON schema enforcement. It is released under the Mistral Ecosystem Open-Weight License. A quantized NVFP4 checkpoint is available for Blackwell GPUs. It scores 48% on the Artificial Analysis Coding Agent Index. It scores 26 on the Artificial Analysis Intelligence Index v5. It ships with embedded DSpark speculative decoding for up to 1.9x throughput.
What changed
Previous Mistral models had fewer parameters, shorter context windows, and no native multimodal input.
What it unlocks
Running a single model for instruction following, reasoning, coding, and agentic workflows with long context and multimodal input.
- ~1T total parameters
- ~52B active parameters per token
- 128 routed experts
- 1M context length
What you need to act on it
- NVIDIA Grace Blackwell GPUs for optimal performance
- vLLM, SGLang, or Transformers for local serving
Sources