Mistral Large 4 Released on Hugging Face
Mistral has introduced its new multimodal, multilingual Mixture-of-Experts model, Mistral Large 4, which arrives on Hugging Face with an expansive context window and powerful capabilities tailored for cybersecurity and coding agents.
Radwan AbdulhadiEditor and technology writer
Introduction to Mistral Large 4
Mistral has officially launched its latest multimodal, multilingual Mixture-of-Experts (MoE) artificial intelligence model, Mistral Large 4, also known as Le Chonk. The release is detailed extensively on the Hugging Face Blog, highlighting the model's architecture, capabilities, and benchmarks. Designed to accept both text and images as inputs while generating text, the model scales up significantly in parameter count and structural complexity compared to previous iterations. Further details are available from Hugging Face Blog in the original source material.
According to technical documentation, Mistral Large 4 features approximately 1 trillion total parameters, with 49 billion active parameters per token, and incorporates a massive 1-million-token context window. The model is distributed under the Mistral Ecosystem Open-Weight License, an MIT derivative that establishes specific limitations for model-as-a-service offerings, while community-created GGUFs have also been developed by Unsloth and the llama.cpp team.
Architecture and Technical Specifications
The underlying architecture of Mistral Large 4 combines a large pool of experts with sparse attention mechanisms and a 1.6-billion-parameter Pixtral vision encoder. Its text decoder comprises 72 layers with a hidden dimension of 6,144 and 64 attention heads. The attention blocks merge Multi-head Latent Attention with a DeepSeek Sparse Attention indexer, where MLA maintains a compressed key-value representation and DSA selects up to 2,048 key positions for each query.
For its feed-forward design, the model utilizes a sparse Mixture-of-Experts configuration across 71 of its 72 decoder layers, where a softmax router selects four out of 128 experts for every token alongside one shared expert, resulting in five active experts per token. Meanwhile, the Pixtral vision encoder contains 30 layers, a hidden dimension of 2,048, and 16 attention heads, processing inputs via 14 by 14 patches and a spatial merge projector.
Cybersecurity and Defensive Capabilities
Mistral Large 4 is positioned heavily for defensive security work, demonstrating standout performance in vulnerability assessment and remediation tasks. The model scores 82 percent on the Artificial Analysis Cyber Index test, which challenges systems to reproduce a real vulnerability in open-source software and subsequently patch it. Mistral notes that several closed frontier models score near zero on the same benchmark because they refuse to attempt the task due to safety filters.
In addition to vulnerability reproduction, internal testing by Mistral showed that the model is useful for malware analysis, vulnerability prioritization, and detection-rule generation. Furthermore, the system solves 93 percent of a 40-challenge suite drawn from security competitions via Cybench, and demonstrates 93.3 percent attack resistance on Lakera's B3 benchmark against prompt-injection attacks.
Coding, Benchmarks, and Ecosystem Support
Beyond cybersecurity, Mistral Large 4 delivers strong benchmark results across coding agent tasks and professional workflows. On coding evaluations, it achieves 61.7 percent on DeepSWE v1.1, 59.4 percent on SWE-Atlas-QnA, and 28.3 percent on Terminal-Bench 4. Its combined Coding Agent Index score places it competitively within the open-weight ecosystem, supported further by scores of 59.9 percent on AutomationBench and 1,393 Elo on AA-Briefcase.
The model is shipped with day-0 inference support across popular frameworks. Developers can utilize Transformers, vLLM, and llama.cpp for local or scaled execution. Remote and serverless deployment options are also available through dedicated Inference Endpoints and Inference Providers, enabling flexible integration for community builders and enterprise environments alike.
Sources
- Hugging Face BlogWelcome Le Chonk: Mistral Large 4 on Hugging Face
Continue chronologically
Related stories
Hugging Face Transformers Now Supports llama.cpp GGUF Models
Hugging Face has added native support for GGUF model files to its Transformers library, enabling developers to run quantized AI locally.
Mistral Releases Trillion-Parameter AI Model Weights
French artificial intelligence company Mistral has released the weights for its trillion-parameter open-source model, targeting cybersecurity.
Mistral Says Le Chonk Leapfrogs China in AI Race
French AI lab Mistral claims its refined open-weight model Le Chonk matches or outperforms leading systems from DeepSeek and Z.ai in coding and cybersecurity.
IBM Announces Granite Speech 5.0 Collection
IBM unveils Granite Speech 5.0, introducing a new speech-language model, a faster multilingual Turbo CTC, and an updated 2B Plus model for enterprise audio.
