Mistral Large 4 (Le Chonk) release notes

TL;DR

Mistral Large 4 (aka Le Chonk), a 1 trillion‑parameter multimodal Mixture‑of‑Experts model with a 1 M‑token context window, is released on Hugging Face in FP8 and NVFP4 formats, offering strong performance for cybersecurity, coding agents, and vision‑language applications while remaining open‑weight under the Mistral Ecosystem Open‑Weight License.

Model Overview

Mistral Large 4 is a multimodal, multilingual MoE model that accepts text and images and outputs text. It contains roughly 1 trillion total parameters, of which 49 billion are active per token. The model supports a 1 million‑token context window, enabling extremely long‑range reasoning.

Two official checkpoint formats are provided:

  1. FP8 – standard 8‑bit floating‑point checkpoint (mistralai/Mistral-Large-4-1T-A52B).
  2. NVFP4 – a variant optimized for NVIDIA Blackwell GPUs (mistralai/Mistral-Large-4-1T-A52B-NVFP4).

Community‑generated GGUF quantizations are also available via the Unsloth and llama.cpp teams.

The model is distributed under the Mistral Ecosystem Open‑Weight License, an MIT‑derived license that restricts model‑as‑a‑service deployments.

Key Technical Features

Sparse Architecture

  • Mixture‑of‑Experts (MoE): 72‑layer decoder, 128 experts per MoE layer, with a softmax router selecting four experts per token plus one shared expert (5 active experts total).
  • Sparse Attention: Combines Multi‑head Latent Attention (MLA) with DeepSeek Sparse Attention (DSA) indexer, selecting up to 2,048 key positions per query.
  • RoPE Variant: Uses LeRope, a YaRN‑scaled rotary embedding with a learned multiplier per frequency.
  • Layer Pattern: 3‑to‑1 RoPE‑to‑NoPE pattern (three layers use RoPE, every fourth uses NoPE) and shared KV state across layers.

Vision Encoder

  • Pixtral encoder: 1.6 B parameters, 30 layers, hidden size 2,048, 16 attention heads.
  • Processes images as 14 × 14 patches with a patch‑merge projector (spatial merge size 2).

Performance Highlights

Cybersecurity

  • 82 % on the Artificial Analysis Cyber Index (reproduce‑then‑patch task), outperforming closed‑source frontier models that often refuse the task.
  • Strong internal results on malware analysis, vulnerability prioritisation, and detection‑rule generation.
  • High refusal rates on harmful prompts (JailbreakBench, StrongREJECT, AgentHarm), indicating robust safety behaviour.

Coding & Agentic Workflows

  • 61.7 % on DeepSWE v1.1, 59.4 % on SWE‑Atlas‑QnA, 28.3 % on Terminal‑Bench 4.
  • Combined Coding Agent Index score of 49.8 %, surpassing DeepSeek V4 Pro 0813 and Qwen 3.8 Max.
  • Business‑workflow scores: 59.9 % on AutomationBench and 1,393 Elo on AA‑Briefcase.

General Multimodal Benchmarks

  • 93.3 % attack resistance on Lakera’s B3 benchmark.
  • Ranked 5th among open‑weight models on the Artificial Analysis Intelligence index (as of 29 Oct 2026).

Inference Ecosystem

Transformers

  • AutoModelForMultimodalLM with AutoProcessor.
  • Supports three reasoning‑effort levels (low, medium, high); default is high.
  • Example script provided for text‑only and image‑text prompts. Distributed launch via torchrun -nproc-per-node 8.

vLLM

  • NVFP4 checkpoint offers reduced memory footprint on NVIDIA Blackwell GPUs.
  • Includes DSpark drafter for speculative decoding.
  • Full launch command and simple curl example included.

llama.cpp

  • GGUF quantizations (Q2_K, IQ2_XXS, Q4_K_M, NVFP4, Q8_0, BF16).
  • Single‑Apple‑Silicon‑Ultra (M3 Ultra) run achieves ~90 t/s prefill and 15 t/s generation for Q2_K.
  • Multi‑GPU performance reported for RTX6000 Pro and NVFP4 configurations.

Fine‑Tuning Support

Day‑0 fine‑tuning is available through TRL (TensorFlow Reinforcement Learning) and PEFT. The provided script fine‑tunes the FP8 variant on the Multilingual‑Thinking dataset using LoRA adapters (r = 16, α = 32) on an 8 × B300s node.

Deployment Options

Hugging Face Inference Endpoints

  • Deployable recipe for 8 × NVIDIA H200 GPUs using vLLM with tensor and expert parallelism.
  • Handles up to 131 K‑token context, 16 concurrent requests, ~0.6 s time‑to‑first‑token for 1 K‑token prompts, and up to 138 t/s generation with DSpark drafts.
  • CLI command: hf endpoints catalog deploy --repo mistralai/Mistral-Large-4-1T-A52B --name mistral-large-4.

Dell Enterprise Hub

  • Optimized Docker recipe for Dell AI servers (e.g., NVIDIA B300) via the dell-ai CLI.
  • Example deployment command provided.

Community Resources

  • GGUF quantizations from Unsloth and llama.cpp are hosted on the Hugging Face Hub.
  • Interview with Pierre Stock on the model’s cyber capabilities is available on YouTube.
  • Users are encouraged to share fine‑tuned models, quantizations, and downstream artifacts on the Hub.

Implications

  • Open‑weight security tooling: Researchers and enterprises can build custom defensive security pipelines without reliance on closed APIs, enabling reproducible vulnerability analysis and incident response.
  • Scalable multimodal reasoning: The 1 M‑token context and sparse attention design make the model suitable for long‑document summarisation, code‑base analysis, and vision‑language tasks.
  • Ecosystem integration: Immediate support across Transformers, vLLM, and llama.cpp lowers the barrier to adoption for both research and production workloads.

Quick Start Checklist

  1. Choose a checkpoint (FP8 or NVFP4) from the Mistral Large 4 collection.
  2. Install the required libraries (transformers, vllm, llama.cpp, trl, peft).
  3. Run the provided inference script or launch a vLLM server.
  4. (Optional) Fine‑tune with TRL on your domain data.
  5. Deploy via Hugging Face Inference Endpoints or Dell Enterprise Hub for production workloads.

Le Chonk is now ready for experimentation and production. Share your results, fine‑tunes, and quantizations on the Hub to accelerate the open‑weight AI ecosystem.

Sources