Mistral Large 4 Arrives on Hugging Face With Cybersecurity Focus

Mistral Large 4 release Hugging Face October 2026

Mistral Large 4, nicknamed Le Chonk, dropped on Hugging Face October 29, 2026. It’s a multimodal, multilingual mixture-of-experts model that takes text and images and spits out text. The release packs roughly one trillion total parameters with 49 billion active at any given moment, plus a context window stretching to one million tokens. Mistral ships two official precision variants, FP8 and NVFP4, while the Unsloth and llama.cpp crowds have already churned out GGUF quantizations. The weights land under the Mistral Ecosystem Open-Weight License, an MIT derivative that tacks on restrictions for model-as-a-service deployments.

Cybersecurity focus defensive security use cases

The model's headline grab centers on cybersecurity. Mistral reports an 82 percent score on the Artificial Analysis Cyber Index, a test asking a model to reproduce a real vulnerability in open-source software and then patch it. The company says several closed frontier models score near zero on the same benchmark simply because they refuse to engage. Mistral argues this capability matters for defenders since fixing a flaw usually requires reproducing it first. Internal testing also showed promise for malware analysis, vulnerability prioritization, and writing detection rules-tasks the model wasn't explicitly trained to handle. On Cybench, a suite of forty challenges drawn from security competitions, the model solves 93 percent. It also hits 93.3 percent attack resistance on Lakera's B3 benchmark for prompt-injection robustness.

Mistral positions the model primarily for defensive security work. The company notes that vulnerability research and incident response often look like attacks in model inputs, which can trip refusal filters in closed systems. Teams relying on an API can also lose access mid-incident. Open weights put both the usage policy and the deployment under the user's control. At the same time, the model still declines harmful requests. On cyber prompts from JailbreakBench, StrongREJECT, and AgentHarm, its average refusal rate exceeds that of any other open-weight model.

For builders, the open-weight release unlocks downstream use cases that would be tough with a closed API. A team could construct a custom cybersecurity system around its own code base, crash logs, and malware samples. Using TRL, developers can fine-tune smaller models from Mistral Large 4 traces that patch their specific code. Because reproduce-then-patch tasks carry a verifiable reward, they fit neatly into reinforcement-learning environments that could further train specialized security models. Open weights also let researchers measure refusal behavior directly instead of trusting a provider's description.

Model architecture specifications text vision encoders

Architecture details reveal a 72-layer text decoder with a hidden dimension of 6,144 and 64 attention heads. The attention blocks combine multi-head latent attention with a DeepSeek sparse attention indexer that selects up to 2,048 key positions for each query. A 3:1 RoPE-to-NoPE pattern runs through the layers, meaning three layers use rotary position embeddings and every fourth layer uses no position embeddings. The model shares key-value state across layers. The first decoder layer employs a dense feed-forward network while the remaining 71 layers use mixture-of-experts blocks. A softmax router picks four of 128 experts per token, and one shared expert always runs alongside them for five active experts per token. Mistral uses LeRope, a YaRN-scaled rotary embedding with one learned multiplier per frequency.

The vision side relies on a 1.6 billion parameter Pixtral encoder with 30 layers, a hidden dimension of 2,048, and 16 attention heads. It processes 14 by 14 patches and uses a patch-merge projector with a spatial merge size of two.

Benchmarks inference support transformers vLLM llama.cpp

Benchmark results published by Mistral show the model ranking fifth among open-weight models on the Artificial Analysis Intelligence Index as of the release date. On coding, it scores 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 of 49.8 percent places it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. For knowledge-work automation, it reaches 59.9 percent on AutomationBench and 1,393 Elo on AA-Briefcase, which evaluate business workflows and long-horizon professional tasks.

Day-zero inference support covers transformers, vLLM, and llama.cpp. The transformers implementation provides AutoModelForMultimodalLM with AutoProcessor handling conversation and image preparation. The model accepts low, medium, or high reasoning effort, defaulting to high, and reasoning cannot be disabled. Remote deployment options include Hugging Face Inference Endpoints and serverless inference through Inference Providers. Fine-tuning is supported via TRL, and the model is also available through the Dell Enterprise Hub.

Developers who need strong cybersecurity reasoning in an open-weight package should evaluate Mistral Large 4 against their existing tooling. The combination of high vulnerability reproduction scores, prompt-injection resistance, and open deployment flexibility makes it a candidate for internal security automation, especially where API dependencies or content filters have been a blocker. Teams building coding agents may also want to test it given the reported lead on the Coding Agent Index. As with any new model, running targeted evaluations on proprietary data before production use remains the prudent step.

Frequently asked questions

When was Mistral Large 4 released and where can I find it?

Mistral Large 4, nicknamed Le Chonk, was released on Hugging Face on October 29, 2026. The model weights are available under the Mistral Ecosystem Open-Weight License, an MIT derivative with restrictions on model-as-a-service deployments.

What are the key architecture specifications of Mistral Large 4?

Mistral Large 4 is a multimodal mixture-of-experts model with roughly one trillion total parameters and 49 billion active parameters. It features a 72-layer text decoder with 6,144 hidden dimensions, 64 attention heads, a 3:1 RoPE-to-NoPE pattern, and a 1.6B parameter Pixtral vision encoder. The context window supports up to one million tokens.

How does Mistral Large 4 perform on cybersecurity benchmarks?

The model scores 82% on the Artificial Analysis Cyber Index for vulnerability reproduction and patching, solves 93% of Cybench's 40 security challenges, and achieves 93.3% attack resistance on Lakera's B3 prompt-injection benchmark. It also shows promise for malware analysis, vulnerability prioritization, and detection rule writing without explicit training on these tasks.

What license governs Mistral Large 4 and what are the deployment implications?

The model uses the Mistral Ecosystem Open-Weight License, an MIT derivative that adds restrictions for model-as-a-service deployments. Open weights give users control over both usage policy and deployment, avoiding API dependency risks during incidents and allowing custom fine-tuning on proprietary code bases and security data.

What inference frameworks and deployment options support Mistral Large 4 at launch?

Day-zero inference support includes transformers (AutoModelForMultimodalLM with AutoProcessor), vLLM, and llama.cpp. Remote deployment options cover Hugging Face Inference Endpoints and serverless inference through Inference Providers. Fine-tuning is supported via TRL, and the model is also available through the Dell Enterprise Hub.

Source: Hugging Face