Mistral Large 4 (“Le Chonk”) Arrives on Hugging Face with Multimodal Open‑Weight Capabilities

What Mistral Large 4 Offers

Mistral Large 4, nicknamed Le Chonk, is a multimodal, multilingual mixture‑of‑experts model that accepts text and images and returns generated text.

The model totals roughly one trillion parameters but activates only 49 billion at inference time.

It supports an unprecedented context window of one million tokens.

Two official precision variants are released: FP8 and NVFP4, and community‑built GGUF formats are available from Unsloth and the llama.cpp team.

Distribution follows the Mistral Ecosystem Open‑Weight License, an MIT‑style license that restricts model‑as‑a‑service deployments.

Architecturally, the model blends a large pool of expert feed‑forward networks with sparse attention driven by a Pixtral vision encoder.

For each token, the language decoder selects a small subset of experts, while an attention indexer chooses which prior token positions to attend to.

This dual sparsity reduces compute while preserving performance across text and vision inputs.

All checkpoints are hosted in the Mistral Large 4 collection on Hugging Face.

Cybersecurity Strengths

On the Artificial Analysis Cyber Index, which requires reproducing a real open‑source vulnerability and then patching it, Mistral Large 4 achieves an 82 percent success rate.

By contrast, several closed‑frontier models score near zero because they refuse to attempt the task.

Internal tests also show usefulness for malware analysis, vulnerability prioritisation, and authoring detection rules, despite no explicit training on those tasks.

The model’s open‑weight nature lets defenders run the system on private codebases, crash logs, and malware samples without relying on external APIs that might be throttled during an incident.

When presented with malicious prompts from benchmarks such as JailbreakBench, StrongREJECT, and AgentHarm, the model’s average refusal rate exceeds that of any other open‑weight model.

These refusal behaviours can be inspected directly by researchers, offering transparent insight into safety performance.

Fine‑tuning pipelines like TRL can distill smaller models from Mistral Large 4 traces, enabling reproducible‑then‑patch reward signals for reinforcement‑learning security agents.

Open‑Weight Ecosystem and Deployment

Support for inference spans Transformers, vLLM, and llama.cpp, giving developers flexibility across hardware stacks.

Fine‑tuning is facilitated through the TRL library, while Dell Enterprise Hub provides managed inference endpoints for enterprise workloads.

Users can experiment instantly via the “Try it out” widget on the model’s Hugging Face page.

By releasing the weights openly, Mistral empowers the community to build bespoke security tools, measure refusal rates, and iterate on model behaviour without vendor lock‑in.

An interview with Pierre Stock, published alongside the release, delves deeper into the model’s defensive security applications.

Why This Matters