Description
Mistral Large 4, also known as "Le Chonk," represents Mistral's most advanced and capable model to date. It is a 1 trillion-parameter natively multimodal Mixture-of-Experts model, featuring 49 billion active parameters per token (52 billion including embeddings and output layers). The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs within Mistral's European datacenters, ensuring robust performance and scalability.
Read More
This generalist model unifies instruction following, reasoning, coding, and agentic workflows into a single set of weights. While versatile across many domains, Mistral Large 4 demonstrates exceptional performance in specialized areas such as cybersecurity, finance, and law. It also surpasses frontier closed models in visual grounding capabilities, making it highly effective for tasks involving image analysis.
A key strength of Mistral Large 4 is its native fluency in over 160 languages, encompassing all official languages of the European Union. This extensive language support makes it a powerful tool for global applications. The model features a substantial 1M context length, enabling it to handle long-context workloads efficiently. Its multimodal input capability allows it to accept both text and image inputs, generating text outputs.
Architecturally, Mistral Large 4 utilizes a granular Mixture-of-Experts design with 128 routed experts, where 4 experts are active per token, alongside 1 shared expert. It also includes an embedded DSpark speculative decoding draft head, which can significantly speed up decoding without altering the target model's outputs, offering up to 1.9x higher throughput in certain scenarios. A quantized NVFP4 checkpoint is also available, providing a smaller memory footprint, faster inference, and equivalent performance, specifically targeting Blackwell GPUs and supporting vLLM and SGLang.
Mistral Large 4 is released under the Mistral Ecosystem Open-Weight License, a modified MIT license that permits both commercial and non-commercial use. However, operating a Model-as-a-Service business with Mistral Large 4 requires a separate commercial license from Mistral. The model is available on the Mistral API as `mistral-large-4` and can be deployed and fine-tuned using various libraries such as vLLM, SGLang, Transformers, llama.cpp, Axolotl, Unsloth, and TRL.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
Mistral Large 4 Highlights
1 Trillion Total Parameters
52 Billion Active Parameters per Token
Natively Multimodal Input (Text and Image)
1 Million Context Length
Multilingual Support (160+ languages)
Configurable Reasoning Mode (Low, Medium, High)
Advanced Vision Capabilities for Image Analysis
Strong System Prompt Adherence
Native Function Calling and JSON Output
Strict JSON Schema Enforcement
State-of-the-Art Cybersecurity Performance
Embedded DSpark Speculative Decoding Draft Head
Quantized NVFP4 Checkpoint Available
Open-Weight License for Commercial and Non-Commercial Use
Getting Started with Mistral Large 4
Access Model: Find Mistral Large 4 on Hugging Face or via the Mistral API.
Deploy with vLLM: Use the official vLLM Docker image for production-ready inference, configuring parameters like `--tool-call-parser mistral` and `--reasoning-parser mistral`.
Deploy with SGLang: Utilize the official SGLang Docker image, enabling speculative decoding with `--speculative-algorithm DSPARK`.
Deploy with Transformers: Install Transformers from main and use the provided Python inference snippet for deployment on multiple GPUs.
Fine-tune Model: Use libraries like Axolotl, Unsloth, or TRL for fine-tuning Mistral Large 4.
Integrate with Mistral Vibe: Install Mistral Vibe and configure it to use Mistral Large 4 via the Mistral API or a local vLLM server.
Configure Reasoning Effort: Set `reasoning_effort` to 'low', 'medium', or 'high' based on prompt complexity.
Utilize System Prompts: Define clear environments and use cases with system prompts for effective tool leveraging.
Mistral Large 4's Use Cases
- Complex Agentic Workflows
- Cybersecurity Analysis
- Financial Modeling & Analysis
- Legal Document Processing
- Multilingual Communication
- Long-Context Information Retrieval
- Image and Chart Analysis
- Code Generation & Debugging
- Instruction Following & Automation
- Chatbot & Conversational AI
