AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Welcome Le Chonk: Mistral Large 4 on Hugging Face

Collected Oct 29, 2026

Mistral has released Mistral Large 4, nicknamed Le Chonk, and Hugging Face has published a walkthrough of the model's architecture, benchmarks, inference paths and fine-tuning support. It is a multimodal, multilingual Mixture-of-Experts model from Mistral that accepts text and images and generates text. The headline numbers are large: about 1 trillion total parameters with 49 billion active per token, and a context window of 1 million tokens.

The model ships in two official variants, FP8 and NVFP4, with GGUFs produced by Unsloth and the llama.cpp team. It is distributed under the Mistral Ecosystem Open-Weight License, described as an MIT derivative that adds restrictions on model-as-a-service offerings. Mistral is positioning it for cybersecurity work, coding agents and multimodal tasks.

The architecture combines two kinds of sparsity. The text decoder has 72 layers, a hidden dimension of 6,144 and 64 attention heads. Attention pairs Multi-head Latent Attention, which keeps a compressed key-value representation, with a DeepSeek Sparse Attention indexer that picks up to 2,048 key positions per query — one mechanism selects which feed-forward networks run, the other selects which context positions are attended to. Layers follow a 3:1 pattern of rotary position embeddings (RoPE) to NoPE, and KV state is shared across layers. Only the first decoder layer uses a dense feed-forward network; the other 71 are MoE blocks where a softmax router picks four of 128 experts per token, joined by one always-on shared expert, for five active experts per token. Position encoding uses LeRope, a YaRN-scaled RoPE with one learned multiplier per frequency. The vision side is a 1.6B-parameter Pixtral encoder with 30 layers, hidden dimension 2,048 and 16 attention heads, using 14 x 14 patches and a patch-merge projector with spatial merge size 2.

Benchmarks are strong but varied. On coding, the model scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4; its combined Coding Agent Index of 49.8% is reported ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. On knowledge-work evaluations it reaches 59.9% on AutomationBench and 1,393 Elo on AA-Briefcase. As of October 29, 2026, it is the fifth-best open-weights model on the Artificial Analysis Intelligence index.

The cybersecurity results are the most unusual part. Mistral Large 4 scores 82% on the Artificial Analysis Cyber Index, the highest reported score on a test that asks a model to reproduce a real vulnerability in open-source software and then patch it. Mistral says several closed frontier models score near zero on the same test because they refuse to attempt the work, and that the model was also useful in internal testing for malware analysis, vulnerability prioritisation and writing detection rules without being explicitly trained for those. It solves 93% of 40 Cybench security-competition challenges and shows 93.3% attack resistance on Lakera's B3 prompt-injection benchmark. It still refuses harmful requests: on cyber prompts from JailbreakBench, StrongREJECT and AgentHarm its average refusal rate is higher than any other open-weight model, according to Mistral.

The reasoning here is worth spelling out. Vulnerability research and incident response can look like attacks in model input, so provider-side filters may block legitimate defensive work, and an API dependent on a third party can be cut off mid-incident. Open weights move both usage policy and deployment under the operator's control. A likely trade-off is that a model willing to work through exploit reproduction demands more governance from whoever deploys it; the licence's model-as-a-service limits are one lever there.

Tooling support is broad on day zero. Transformers offers AutoModelForMultimodalLM with AutoProcessor handling conversation and image inputs, plus a reasoning_effort setting of low, medium or high (default high). Notably, reasoning cannot be turned off — passing none still yields high effort. vLLM supports the NVFP4 checkpoint, which needs NVIDIA Blackwell GPUs and less memory, and pairs with a DSpark drafter for speculative decoding speed-ups. llama.cpp and Llama App support the model, with quants IQ2_XXS, Q2_K, Q4_K_M, NVFP4, Q8_0 and BF16. The 2-bit quants run on a single Apple Silicon Ultra machine with 512 GB of unified RAM; others need distributed inference. On one M3 Ultra, Q2_K reached roughly 90 tokens per second prefill and 15 tok/s generation. On four RTX6000 Pro cards, NVFP4 at 32K context prefilled at 130-590 t/s and generated around 10 t/s, while IQ2_XXS at 800K context managed 500 t/s prefill and 30 t/s generation at short lengths but dropped to 129 t/s and 3 t/s at long context.

Fine-tuning is available through TRL with an upgrade of transformers, trl and peft. The example script trains multilingual reasoning on the FP8 variant across eight B300 GPUs using LoRA, FSDP and expert parallelism. On the hosted side, an Inference Endpoints recipe runs vLLM with FP8 weights across eight NVIDIA H200s, serving up to 131K tokens of context and 16 concurrent requests; reported figures include about 0.6 seconds to first token for a 1K-token prompt, roughly 65 tok/s generation (120-138 tok/s when DSpark drafts are accepted), about 4,300 tok/s aggregate prefill on four simultaneous 16K prompts, and around 480 tok/s aggregate generation across 1,000 ShareGPT conversations. Dell Enterprise Hub offers a dell-ai CLI recipe for H200, B200, GB200 and GB300 servers.

Why it matters: teams weighing open-weight models for agentic coding or security work now have a roughly 1T-parameter option with a 1M-token context, day-zero support across Transformers, vLLM, llama.cpp and TRL, and an explicit license carve-out for model-as-a-service. The practical constraint is hardware — full-precision deployment needs multi-GPU nodes, and aggressive quantization is what brings it to single machines, with the long-context throughput collapse a likely cost. My reading is that the security positioning is the differentiating bet: if defenders can run and fine-tune a capable model on their own vulnerability data without provider filters, that changes incident-response workflows, though it also shifts safety obligations onto the deployer.

Read at Hugging Face Blog

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt