Mistral released Mistral Large 4 on Thursday, launching a multimodal mixture-of-experts model with roughly one trillion total parameters and 49 billion active parameters. Hugging Face reported day-zero platform availability for the release, which is distributed under the Mistral Ecosystem Open-Weight License, an MIT derivative that restricts model-as-a-service offerings. The model takes text and image inputs and operates across a context window of one million tokens.
Official release packages include FP8 and NVFP4 checkpoints alongside GGUF quantizations from Unsloth and the llama.cpp project. The text decoder contains 72 layers, a hidden dimension of 6,144, and 64 attention heads. Its attention mechanism integrates Multi-head Latent Attention with a DeepSeek Sparse Attention indexer that selects up to 2,048 key positions per query. Every fourth decoder layer substitutes RoPE for NoPE, using YaRN-scaled LeRope across the remaining blocks. The first layer uses a dense feed-forward network, while the next 71 layers route each token to four of 128 experts alongside one shared expert. A 1.6-billion-parameter Pixtral vision encoder processes visual inputs across 30 layers.
Benchmark Performance
Artificial Analysis ranked Mistral Large 4 as the fifth-best open-weight model on its Intelligence Index as of October 29, 2026. On the Artificial Analysis Cyber Index, the model recorded an 82% score on tasks requiring it to reproduce and patch real-world open-source vulnerabilities. Hugging Face noted that closed frontier models scored near zero on that evaluation due to safety filter refusals. Mistral Large 4 also solved 93% of 40 security tasks on Cybench and showed 93.3% attack resistance on Lakera's B3 prompt-injection benchmark.
Software evaluation benchmarks yielded a 61.7% score on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4. Its combined Coding Agent Index score of 49.8% placed it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. On enterprise evaluation suites, the model scored 59.9% on AutomationBench and reached 1,393 Elo on AA-Briefcase.
Hardware and Deployment
Execution frameworks including vLLM, llama.cpp, and Transformers support the model natively. The implementation uses a reasoning effort parameter set by default to high, which cannot be disabled. On local hardware, 2-bit quantizations run on an Apple Silicon Ultra system with 512 gigabytes of unified memory, generating approximately 15 tokens per second on an M3 Ultra processor.
Dell enterprise servers deploy the NVFP4 model format via dedicated tooling on H200, B200, GB200, and B300 hardware configurations. Hugging Face also provides managed hosting across eight Nvidia H200 accelerators. That setup uses an integrated DSpark draft head for speculative decoding, generating between 120 and 138 tokens per second on chat prompts.
