How to Run Mistral Large 4 (1T-A52B) Locally: The Hardware It Actually Needs

Mistral Large 4 is Mistral AI's largest release yet: a 1 trillion parameter multimodal mixture-of-experts model with 49 billion active parameters per token, a 1 million token context window, and FP8 weights split across 54 safetensors shards. Mistral's own vLLM example asks for 8 GPUs, so this is a datacenter model, not something a single workstation can load.

Written by Priya Raghunathan, local-AI hardware reviewer. I check parameter counts, licenses, hardware requirements and install commands against the primary source before recommending a setup path.

A new top-20 entrant on Hugging Face's overall trending list

Mistral Large 4 entered the top 20 of Hugging Face's overall trending list about a month after Mistral published it on September 30, 2026. When this entry was filed, the model page shows 1,291 likes and 1,239 downloads, and the Mistral Vibe GitHub repo that ships alongside it has 5,245 stars. The most-liked linked demo is ml4-le-chonk-speaks-160-languages (3 likes), a browser demo of the model's multilingual voice features.

Specs at a glance

What it actually is

Mistral Large 4 is a single checkpoint meant to cover instruction following, reasoning, coding and agentic tool use, rather than separate models for each. It takes text and images as input and answers in text, with a vision encoder built to handle technical drawings, charts and documents at varied sizes and aspect ratios. A reasoning_effort setting of low, medium or high lets you trade speed for depth on a per-request basis, and Mistral reports it is natively fluent in more than 160 languages, including every official EU language.

The checkpoint also embeds a DSpark speculative decoding draft head: three small dense layers that propose blocks of three tokens ahead of the main model, which Mistral says can push throughput up to 1.9 times higher than serving the full model alone. It's on by default in the SGLang example below. For vLLM it's an opt-in flag, shown right after the base serving command.

Hardware: what 1 trillion parameters actually costs you

Mistral ships the main checkpoint as 54 safetensors shards in FP8, and both the official vLLM and SGLang serving commands set a tensor-parallel size of 8. That's the clearest signal of what this model expects: eight GPUs working together, not one. Mistral also publishes a quantized NVFP4 build with a smaller footprint, but it's still built for Blackwell-class datacenter cards (B200, B300, GB200, GB300), not a single desktop GPU of any size on the market today.

Benchmarks

Mistral measured Large 4 against other open-weight models on two of Artificial Analysis's public indexes. On the Coding Agent Index, it scores 48%, calling it "effectively tied" with GLM-5.3, the leading open-weight model on that benchmark, while using 47% fewer tokens per task on average. On the broader Intelligence Index v5, it scores 26, which Mistral describes as the highest among western open-weight models, though it still trails the top-scoring open models out of China on that aggregate measure.

Run it with vLLM

Mistral publishes a ready-made Docker image, so there's no separate wheel install to get the serving stack right:

DSpark isn't on by default here. To turn it on, add --speculative-config '{"method": "dspark", "num_speculative_tokens": 3}' to the command above.

If you're short on GPU memory, Mistral's own docs say to lower --max-model-len or --max-num-seqs before touching anything else. Once the server is up, it speaks the OpenAI chat completions API, so any existing OpenAI client code works against it by pointing base_url at your own host instead of api.openai.com.

Mistral Vibe: the coding agent built on top of it

Mistral Vibe is Mistral's own terminal coding agent, and it's the fastest way to point your own vLLM server at real work instead of a single chat prompt. Install it with a one-line script, then add your server as a local provider:

From there, edit ~/.vibe/config.toml to add a provider pointing at your server's api_base (for example http://your-host:8000/v1), set active_model to that provider's alias, and Vibe will drive Mistral Large 4 through the same OpenAI-compatible endpoint the vLLM server exposes.

FAQ

What is Mistral Large 4 used for?

General-purpose instruction following, reasoning, coding and agentic tool use, with text and image input. Mistral's own benchmarks highlight it specifically for agentic coding tasks, cybersecurity, finance and legal work.

How many parameters does it actually have?

About 1 trillion total, with 49 billion active per token during inference (52 billion once you include embeddings and output layers). It's a mixture-of-experts model, so only a fraction of the total weights do work on any given token, but all of them still have to be loaded in memory.

Can I run it on a single GPU?

No. Mistral's own vLLM and SGLang serving examples both set a tensor-parallel size of 8, and even the smaller NVFP4 quantized build targets multi-GPU Blackwell-class datacenter hardware. There is no published quantization that gets this model onto one consumer card.

Is it free to use commercially?

The weights are licensed for commercial and non-commercial use under Mistral's own Ecosystem Open-Weight License, but running the model as the backend of a paid model-as-a-service product requires a separate commercial license from Mistral directly.

Where to go from here

• Not ready for an 8-GPU build? Run DeepSeek V4 locally: which size actually fits your hardware walks through picking a mixture-of-experts size that matches what you actually own.
• Want to compare it against other open models first? Browse the local AI model directory to see specs side by side before you commit to a download or a rental.