Mistral Large 4 Weights Arrive on Hugging Face: About 1T-Parameter MoE, Benchmarks Vendor-Reported

On October 29, 2026, Hugging Face published a post announcing that the weights of Mistral Large 4, nicknamed Le Chonk, are on the Hub with support in Transformers, vLLM and llama.cpp. It is a multimodal, multilingual Mixture-of-Experts (MoE) model that takes text and images and produces text, with about 1T total and 49B active parameters and a 1M-token context window, according to the post. Mistral had already announced the model as a public preview on October 6 and promised weights by the end of the month, so October 29 is the weight release, not the first appearance.

Two parameter counts in the sources need to be kept apart. The Hugging Face post says 49B active, yet the repository name carries A52B. The model card explains that 49B parameters are active per token, or 52B including the embedding and output layers, and Mistral's October 6 announcement gives 52B. On licensing, the model ships under the Mistral Ecosystem Open-Weight License, an MIT derivative open to commercial and non-commercial use, except that running a model-as-a-service business needs a separate commercial license from Mistral. There are two official variants, FP8 and NVFP4, plus GGUF quantizations from Unsloth and the llama.cpp team.

On architecture, the text decoder has 72 layers, a hidden size of 6,144 and 64 attention heads. Attention combines Multi-head Latent Attention (MLA) with a DeepSeek Sparse Attention (DSA) indexer that selects up to 2,048 key positions per query. Layers follow a 3:1 pattern in which three use RoPE and every fourth uses NoPE, and KV state is shared across layers. The first layer has a dense feed-forward network, and the other 71 are MoE blocks in which a softmax router picks four of 128 experts per token, plus one shared expert that always runs, for five active experts. The vision side is a 1.6B-parameter Pixtral encoder with 30 layers and 14 by 14 patches.

The benchmark figures are reported by Mistral, and the post labels them that way. For coding: 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, with a combined Coding Agent Index of 49.8% that sits ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. For cybersecurity: a top-five place on the Artificial Analysis Cyber Index, including 82% on the reproduce-a-real-vulnerability-then-patch test, which the post calls the highest score in the index; 93% of the 40 Cybench competition challenges; and 93.3% attack resistance on Lakera's B3. Mistral says several closed frontier models score near zero on that test because they refuse the task. The model still declines harmful requests: on cyber prompts from JailbreakBench, StrongREJECT and AgentHarm its average refusal rate is higher than any other open-weight model. As of October 29 it ranks fifth among open-weight models on the Artificial Analysis Intelligence Index.

On deployment, reasoning effort can be low, medium or high. The default is high and it cannot be turned off, so setting none still runs high. vLLM supports the NVFP4 checkpoint, which needs NVIDIA Blackwell GPUs and ships with a DSpark drafter for speed. llama.cpp offers IQ2_XXS, Q2_K, Q4_K_M, NVFP4, Q8_0 and BF16, and the 2-bit quants run on a single Apple Silicon Ultra machine with 512 GB of unified memory. The authors measured Q2_K on one M3 Ultra at about 90 tokens per second for prefill and 15 for generation, and NVFP4 at 32K context on four RTX 6000 Pro cards at 130 to 590 for prefill and about 10 for generation. On the Inference Endpoints recipe with eight H200 GPUs, a single 1K-token prompt gets its first token in about 0.6 seconds and generates about 65 tokens per second, rising to 120 to 138 on chat prompts where draft tokens are accepted. Under 1,000 ShareGPT conversations the total is about 480 tokens per second. These are Hugging Face measurements on specific setups, not general performance promises.

The post also gives a TRL example that fine-tunes the FP8 model on eight B300 GPUs, and a Dell Enterprise Hub deploy command. Keep three limits in mind: the benchmarks are vendor-reported, open weights do not mean unrestricted commercial use, and a 1T-parameter model needs a lot of memory or several cards even when quantized. Re-test on your own tasks before relying on the scores.