Mistral releases the Le Chonk weights on Hugging Face

A team that cannot send its code to a closed model now has a 1 trillion parameter open option. Mistral AI released the weights for Mistral Large 4 on Hugging Face at 10am Pacific today. The nickname is Le Chonk. The preview was three weeks ago.

Mistral and Hugging Face wrote the release post together. Two official builds are on Hugging Face. One is FP8 and one is NVFP4. GGUF quantizations are also on Hugging Face, from the ggml-org and Unsloth collections. The licence is the Mistral Ecosystem Open-Weight License. Hugging Face calls it an MIT derivative. It limits model-as-a-service offerings. The context window is 1 million tokens.

What a team actually gets

The release has day-one support in Transformers, vLLM, llama.cpp and TRL. Hugging Face lists four ways to deploy it. Transformers is for a local run. vLLM is for a high-performance server. llama.cpp runs the quantized files locally. Hugging Face Inference Endpoints is the managed path.

The hardware bar is high at full precision. The managed endpoint runs on 8 NVIDIA H200 GPUs. The NVFP4 build needs Blackwell GPUs. The smallest community quant, Q2_K, runs on a 512 GB Apple Silicon Mac Studio, Hugging Face says. The model has 128 experts. The router picks 4 per token, plus one shared expert. 49 billion of the 1 trillion parameters are active per token.

The cyber claim stays Mistral's own

The headline numbers are the ones Mistral reported with the preview. They are measured on Mistral's own hardware. 82 percent on the Artificial Analysis Cyber Index. That test asks a model to reproduce a vulnerability in open-source software and patch it. 93 percent on Cybench. Cybench is a set of 40 security-competition tasks. 93.3 percent on Lakera's B3 attack-resistance benchmark. 61.7 percent on DeepSWE v1.1 for software engineering. 59.9 percent on AutomationBench.

Mistral reports that Mistral Large 4 uses 47 percent fewer tokens per task than GLM-5.3, TechCrunch writes. Mistral also says closed frontier models score near zero on the Cyber Index test. They refuse to attempt the work, Mistral says. Every figure here is Mistral's own. A researcher must now re-run each test on the released weights.

What a preview could not give you

Three weeks ago the only way to use Large 4 was Mistral Studio or Mistral's API. A team that must keep model inference on its own hardware had nothing to run. That changed today. The release notes list every quantization. The fine-tuning path through TRL and LoRA is in the post. A reader can now measure the refusal behaviour of Large 4 against a closed competitor directly, Hugging Face says.

Mistral reports managed inference numbers on 8 H200 GPUs. About 0.6 seconds to first token on a 1,000 token prompt. 65 tokens per second generation. 4,300 tokens per second prefill on a 16,000 token prompt. 480 tokens per second under 1,000 ShareGPT conversations in parallel.

In Europe some teams cannot use American or Chinese closed models. For them the self-hosting answer is stronger today. The size, the hardware bar and the licence still decide whether this is the right model for the job. Run the eval on your own prompts first. Do not trust any of Mistral's figures without that.

Source

Welcome Mistral Large 4: Le Chonk, Hugging Face, co-written with Mistral AI. Press coverage: Mistral releases weights for its trillion-parameter AI model, targeting cybersecurity, TechCrunch.