1.05T total / 52B active (49B active transformer parameters per token plus embeddings and output layers; 1.6B vision encoder)
1,048,576 tokens
Mistral Ecosystem Open-Weight License
Final open-weight release of Mistral Large 4 following its October 6 public preview and three additional weeks of reinforcement learning. It is a natively multimodal text-and-image-input, text-output model trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters. The 72-layer granular mixture-of-experts language backbone has 128 routed experts with four selected per token, one shared expert, a 6,144-dimensional hidden state, and a dense first layer. A 30-layer Pixtral-style vision encoder supports variable image dimensions and aspect ratios. The same checkpoint unifies ordinary instruction following and configurable low, medium, or high reasoning, with high as the default. It supports system prompts, native function calling, strict JSON-schema output, visual grounding, and more than 160 languages, including every official European Union language. Mistral reports 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task; 41% on Terminal-Bench 4.0; and 19.7% on Cognition's private FrontierCode 2.0. The model card reports an Artificial Analysis Intelligence Index v5 score of 26, the highest among western open-weight models at release. Other reported results include 51% on the Artificial Analysis Cyber Index, 73.3% on Finch, 84.5% on BrowseComp, and 97.1% attack resistance on the B3 Agent Security Benchmark. These results mix third-party and publisher-run evaluations and should not be treated as interchangeable measures. The checkpoint includes an embedded DSpark speculative-decoding head consisting of three dense MLA draft layers that propose three-token blocks; Mistral reports up to 1.9x higher throughput without changing target-model outputs. The primary FP8 checkpoint contains 54 safetensor shards totaling approximately 1.06 TB. An official Blackwell-targeted NVFP4 release uses NVFP4 MoE experts with FP8 linear layers and contains 32 shards totaling approximately 610 GB; Mistral reports similar evaluation quality with lower memory use and faster inference. Official deployment support includes vLLM, SGLang, Transformers, llama.cpp, and Mistral Vibe. A full technical report is announced but was not yet available at release.