Mistral Large 4 Released: Multimodal MoE Model

At a Glance
| Item | Value |
|---|---|
| Publisher | Hugging Face Blog |
| Published | 2026-10-29 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Mistral AI has released “Mistral Large 4" (also known as Le Chonk), a multimodal, multilingual Mixture-of-Experts (MoE) model. This model accepts text and image inputs and can generate text.
Mistral Large 4 adopts an architecture that combines a large pool of experts with sparse attention and the Pixtral vision encoder. It features two distinct forms of sparsity: the language model selects a small number of experts for each token, while the attention indexer selects past token positions to reference.
Strengths and Use Cases
Mistral Large 4 performs exceptionally well in cybersecurity, coding agents, multimodal processing, and agentic workflows.
In the cybersecurity domain, it is particularly suited for defensive security tasks. It scores 82% in the Artificial Analysis Cyber Index tests, which replicate real-world vulnerabilities in open-source software and apply patches, and solves 93% of the challenges in Cybench, which consists of tasks from security competitions. Furthermore, it is effective for malware analysis, vulnerability prioritization, and detection rule creation. Because it is provided as an open weight, users can control their own usage policies and deployment environments, avoiding the risk of service access being cut off during incident response. It can also be utilized for direct measurement of refusal behavior, building dedicated systems incorporating custom codebases and malware samples, and fine-tuning smaller models using TRL.
For coding and business tasks, it achieves 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4 (a metric measuring whether a model can actually execute commands in a terminal and complete a given task to the end), with a Coding Agent Index score of 49.8%. It also scores 59.9% on AutomationBench, which evaluates business workflows and long-term professional tasks, and 1,393 Elo on AA-Briefcase.
In addition, equipped with a 1.6B parameter Pixtral vision encoder, it supports multimodal tasks combining text and images. It demonstrates 93.3% attack resistance in Lakera’s B3 benchmark, which evaluates robustness against prompt injection attacks.
How to Get It
Mistral Large 4 is distributed under the Mistral Ecosystem Open-Weight License. This license is a derivative of the MIT license, but includes restrictions on providing it as a Model-as-a-Service (MaaS).
As for distribution formats, official FP8 and NVFP4 versions are available. GGUF format quantized models (IQ2_XXS, Q2_K, Q4_K_M, NVFP4, Q8_0, BF16) created by Unsloth and the llama.cpp team are also available.
Supported engines and libraries are as follows:
Transformers Use it by updating the transformers package to the latest version. Process text and image inputs using AutoModelForMultimodalLM and AutoProcessor. Reasoning effort (reasoning_effort) can be selected from low, medium, and high (default).
pip install --upgrade transformers
vLLM Supports NVFP4 checkpoints, which require NVIDIA Blackwell GPUs. DSpark drafters are also supported for acceleration.
llama.cpp / Llama App By installing Llama App, you can start a local server and use it via WebUI or API.
curl -LsSf https://llama.app/install.sh | sh
llama serve -hf ggml-org/Mistral-Large-4-1T-A52B-GGUF:Q2_K
TRL (Fine-tuning) You can perform fine-tuning by combining trl, transformers, and peft.
pip install --upgrade transformers trl peft
Cloud and Enterprise Environments It also supports deployment using Hugging Face Inference Endpoints or the dell-ai CLI in the Dell Enterprise Hub.
pip install dell-ai && dell-ai login
Related Articles
- Meta Releases Muse Spark Open: Multimodal MoE Model
- Phonon: A Lightweight, Fully On-Device Text-to-Speech Model
- Run GGUF Models Directly in Transformers with llama.cpp Support
- BotPsy: 0.8B On-Device Agentic Model by Tether AI Research
What to Read Next
- How to read Terminal-Bench → Benchmark glossary

