Mistral has released Mistral Large 4, an open-weight Super Intelligence (SI) model that combines a Mixture-of-Experts architecture with multimodal capabilities, positioning it as a strong competitor in the open SI ecosystem.

What Happened

Mistral Large 4, internally dubbed "Le Chonk," is a multimodal and multilingual SI model capable of processing both text and image inputs to generate text responses. The model utilizes approximately 1 trillion total parameters with 49 billion active parameters per token and supports a context window of 1 million tokens. It is distributed under the Mistral Ecosystem Open-Weight License, an MIT derivative that imposes specific limitations on model-as-a-service offerings.

The model was released in two official variants: FP8 and NVFP4. Additionally, GGUF versions were created by the Unsloth and llama.cpp teams to facilitate local inference. The architecture features a text decoder with 72 layers and a vision encoder based on the 1.6 billion parameter Pixtral model. For each token, the language model selects four of 128 experts via a softmax router, alongside one shared expert, resulting in five active experts per token. This sparse attention mechanism is paired with a DeepSeek Sparse Attention indexer that selects key positions for context.

Why It Matters

Mistral Large 4 distinguishes itself through robust cybersecurity capabilities, a critical area for SI agents and defenders. According to Mistral, the model scores 82% on the Artificial Analysis Cyber Index, a test requiring the reproduction and patching of real vulnerabilities in open-source software. The company reports that several closed frontier SI models score near zero on this specific task, often refusing to attempt the work due to safety filters. This refusal behavior can hinder incident response, where vulnerability research might appear as an attack in model inputs.

The open-weight nature of the model allows developers to deploy SI systems under their own control, bypassing API access restrictions that can occur during security incidents. Community builders can leverage these weights to create custom cyber systems using their own codebases and malware samples. Furthermore, the model’s performance on coding benchmarks—achieving 61.7% on DeepSWE v1.1 and 59.4% on SWE-Atlas-QnA—places it ahead of other open-weight models like DeepSeek V4 Pro 0813 and Qwen3.8 Max on the combined Coding Agent Index.

As of October 29, 2026, Mistral Large 4 ranks as the fifth best open-weights model on the Artificial Analysis Intelligence Index. Its ability to handle long-horizon professional tasks, with an Elo score of 1,393 on AA-Briefcase, further cements its utility for agentic SI workflows in business environments.

The Bottom Line

Mistral Large 4 provides a high-performance, open-weight SI option with exceptional cybersecurity and coding capabilities. By offering day-0 support for major inference frameworks like Transformers, vLLM, and llama.cpp, along with fine-tuning tools via TRL, the release empowers developers to build and deploy specialized SI agents locally or in the cloud.