Mistral opens a 1T-parameter multimodal model: 82% on reproducing and patching real vulnerabilities

The open-weight ecosystem gains a multimodal Mixture-of-Experts model with about 1T total parameters that takes text and images as input, generates text, and carries a 1M-token context window. Mistral Large 4 was not previously available as open weights; it now ships in two official variants, FP8 and NVFP4, plus GGUFs made by Unsloth and the llama.cpp team, under the Mistral Ecosystem Open-Weight License, an MIT derivative that sets limitations for model-as-a-service offerings. As of October 29, 2026, the model is the fifth best open-weights model on the Artificial Analysis Intelligence index; its combined Coding Agent Index score is 49.8%, which Mistral says places it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max.

The model stacks two forms of sparsity: one selects feed-forward experts in each layer, the other selects context positions in attention. The text decoder has 72 layers, a hidden dimension of 6,144 and 64 attention heads; its attention blocks combine Multi-head Latent Attention, which keeps a compressed key-value representation, with a DeepSeek Sparse Attention indexer that selects up to 2,048 key positions per query. The first decoder layer uses a dense feed-forward network and the other 71 use MoE blocks, where a softmax router selects four of 128 experts per token and one shared expert always runs alongside, for five active experts per token; decoder layers alternate RoPE and NoPE in a 3:1 pattern and share KV state across layers.

Cybersecurity is the capability Mistral emphasizes for this release, and the model is positioned mainly for defensive security work. Mistral reports that several closed frontier models score near zero on the same reproduce-then-patch test because they refuse to attempt the work, while fixing a vulnerability usually requires reproducing it first. On the Artificial Analysis Cyber Index, the model scores 82% on reproducing and patching vulnerabilities, which Mistral calls the highest reported score on that test within the index; it solves 93% of a 40-challenge Cybench suite drawn from security competitions and shows 93.3% attack resistance on Lakera's B3 benchmark.

The report covers the Mistral Large 4 release, its architecture description, benchmark scores and inference support; the benchmark results are published by Mistral and Hugging Face, and the malware analysis, vulnerability prioritisation and detection-rule generation are described as Mistral internal testing.

Mistral's statement that closed frontier models score near zero on the same vulnerability-reproduction test because they refuse to attempt the work comes from Mistral rather than an independent evaluation, and the malware analysis, vulnerability prioritisation and detection-rule generation results also come only from internal testing. In the reported inference configurations, throughput drops markedly at long context: IQ2_XXS at 800K context falls to 129 tok/s prefill and 3 tok/s generation.