Mistral releases open weights for trillion-parameter Large 4
Mistral has released the weights for Mistral Large 4, nicknamed Le Chonk. The model takes text and images as input and outputs text and code. Hugging Face lists about 1 trillion total parameters and 49 billion active parameters, with a context window of 1 million tokens. The Decoder reports 52 billion active parameters. The weights are available on Hugging Face under the Mistral Ecosystem Open-Weight License, an MIT derivative with limits on model-as-a-service offerings.
Mistral is positioning the release mainly for cybersecurity and coding. For teams that run software or manage infrastructure, this matters because open weights mean no provider can cut off access during an incident.
In one line: Mistral Large 4 is an open-weight, trillion-parameter mixture-of-experts model that performs well on cybersecurity and coding benchmarks but still trails other open models on overall intelligence.
Why cybersecurity is the focus
Mistral reports that Large 4 scores 82 percent on the Artificial Analysis Cyber Index task that asks a model to reproduce a real vulnerability in open-source software and then patch it. Hugging Face says several closed frontier models score near zero on the same test because they refuse to attempt the work. The model also shows 93.3 percent attack resistance on Lakera's B3 benchmark and, according to The Decoder, rejects 96.6 percent of malicious cyber requests on average while still handling legitimate defensive work.
The Decoder notes that Mistral calls Large 4 the leading open model for cyber defense, but Mistral's own chart shows two open models ahead: DeepSeek V4.1 Flash at 57 percent and MiMo-V2.6-Pro at 56 percent on the Cyber Index. Large 4 is at 51 percent. This is a real gap between marketing and independent numbers.
Coding position
Hugging Face reports 61.7 percent on DeepSWE v1.1 and 59.4 percent on SWE-Atlas-QnA. The Decoder says the final version reached 41 percent on Terminal-Bench 4.0 after additional reinforcement learning, just ahead of GLM-5.3 at 40 percent. Still, Claude Opus 5.5 tops the Coding Agent Index at 62 points, and Large 4 wins only 22 percent of blind matchups against Opus 5.5. So it is competitive among open models, but not close to the closed frontier on coding.
Architecture and hardware
Large 4 is a mixture-of-experts model. The text decoder has 72 layers, a hidden dimension of 6,144, and 64 attention heads. It uses Multi-head Latent Attention with a DeepSeek Sparse Attention indexer. The vision encoder is a 1.6 billion parameter Pixtral model. Running the standard version requires eight Nvidia B300 GPUs, according to SGLang docs cited by The Decoder. A compressed variant runs on four GB300 GPUs. GGUFs are available from Unsloth and the llama.cpp team, and the model has day-0 support in transformers, vLLM, and llama.cpp.
What it does not yet do
On Artificial Analysis' Intelligence Index version 5, Large 4 scores 26 points in high mode. The Decoder reports DeepSeek V4.1 Flash and GLM-5.3 each score 32 points, and Kimi K3.1 Pro leads the cyber index at 64 percent. Mistral's model is fast at 161 tokens per second and cheaper than some rivals at $0.70 per million input tokens and $2.10 per million output tokens. But it remains behind Chinese open models on broader intelligence tasks. The license also requires a separate agreement for anyone offering Large 4 as a commercial cloud service.
Practical take
For a small team doing client delivery, this is mostly a watch item. Eight B300 GPUs are not something we would buy for a RAG chatbot project. I would run a handful of defensive security and coding prompts through the API to see if the refusal behaviour fits regulated client work. If a client asks for on-prem cyber tooling, Large 4 is now a credible option to test, but only after accounting for hardware, license limits, and the gap between Mistral's cyber claims and the independent numbers.