Mistral releases Mistral Large 4, a 1T-parameter multimodal MoE model with 49B active parameters, on Hugging Face
The open-weight model supports 1M-token context, achieves 82% on a vulnerability reproduction and patching benchmark, and ships with day-0 support for Transformers, vLLM, and llama.cpp.
1 source · cross-referenced
- Mistral Large 4 is a 1 trillion total / 49 billion active parameter multimodal MoE model released under the Mistral Ecosystem Open-Weight License.
Mistral Large 4 (codenamed Le Chonk) is a multimodal, multilingual Mixture-of-Experts model from Mistral that accepts both text and images and generates text. The model exposes roughly 1 trillion total parameters with 49 billion active parameters and supports a context window of up to 1 million tokens. It is distributed under the Mistral Ecosystem Open-Weight License, a derivative of MIT that imposes limits on model-as-a-service deployments.
Cybersecurity is a focal capability: Mistral reports that Mistral Large 4 achieves 82% on the Artificial Analysis Cyber Index’s reproduce-and-patch task, which requires reproducing a real vulnerability in open-source software and then patching it. Mistral states that several closed frontier models score near zero on the same test due to refusal behavior. In internal evaluations, the model also showed utility for malware analysis, vulnerability prioritization, and writing detection rules without explicit training for those tasks.
The model’s refusal behavior was measured on JailbreakBench, StrongREJECT, and AgentHarm, where its average refusal rate exceeded that of other open-weight models. Mistral positions the release to give practitioners control over usage policies and deployment, noting that closed APIs can block legitimate security workflows during incidents.
Architecturally, the text decoder comprises 72 layers with a hidden dimension of 6,144 and 64 attention heads. Mistral combines Multi-head Latent Attention with a DeepSeek Sparse Attention indexer that selects up to 2,048 key positions per query. The model uses a 3:1 RoPE to NoPE pattern across layers and shares KV state across layers. The decoder uses MoE blocks with a softmax router selecting four of 128 experts per token plus one shared expert, yielding five active experts per token.
A 1.6-billion-parameter Pixtral vision encoder handles images, using 14×14 patches and a patch-merge projector with spatial merge size 2. The model ships day-0 support for Transformers, vLLM, and llama.cpp, with GGUF quantizations provided by the llama.cpp team. Remote deployment options include Hugging Face Inference Endpoints and serverless inference via Inference Providers.
- Nov 11, 2026 · Hugging Face
Hugging Face proposes token pooling-aware training to improve compressed multi-vector retrieval
Trust79 - Nov 11, 2026 · TechCrunch — AI
Fractile raises $720M Series C to develop AI chips and compiler stack
Trust79 - Nov 11, 2026 · The Verge — AI
Google’s Tabletop AI experiment lets users build custom collaborative apps via waitlist
Trust79