---
title: "Mistral Large 4 Open Weights: Benchmarks, Price, How to Run It"
description: "Mistral Large 4 is out as open weights: 1T parameters, 49B active, $0.70 in and $2.10 out per million tokens. What it costs, how to run it, and the hardware bill."
canonical_url: https://www.oguzhan.co/mistral-large-4-open-weights/
author: "Oğuzhan Koçaklı"
author_url: https://www.oguzhan.co/about/
date_published: 2026-10-30T08:28:50+00:00
date_modified: 2026-10-30T08:28:53+00:00
language: en
translations:
  tr: https://www.oguzhan.co/tr/mistral-large-4-acik-agirlik/
---

# Mistral Large 4 Open Weights: Benchmarks, Price, How to Run It

I spent this morning reading Mistral’s announcement, the vLLM blog post and recipe, the Hugging Face write-up and the SGLang cookbook page for Mistral Large 4. I have not run the model myself. Every number below comes from those pages, and the benchmark scores are Mistral’s own figures, so read them that way.

My short take: for most readers the API price is the real news. Self-hosting is possible, but it is a job for teams with H200 or B300 class nodes, or teams that need the weights on their own infrastructure for sovereignty reasons. This is not a laptop model.

## What Mistral released

[Mistral’s announcement](https://mistral.ai/news/mistral-large-4-oss/) calls the model “Le Chonk”, and the name fits. Mistral Large 4 is a 1-trillion-parameter model that is natively multimodal, taking text and image input. It went into public preview earlier in October and the weights are now public. Mistral says it was trained from scratch on European infrastructure and that it is “natively fluent in 160+ languages”.

Under the hood it is a mixture of experts. According to the [vLLM blog](https://vllm.ai/blog/2026-10-29-mistral-large-4) and [recipe](https://recipes.vllm.ai/mistralai/Mistral-Large-4-1T-A52B), only 49B parameters are active per token (52B if you count embeddings and the output layer). There are 72 layers, 128 routed experts with top-4 routing plus one shared expert, and a Pixtral vision encoder. It also ships with a built-in DSpark drafter for speculative decoding, which vLLM says gives roughly 2x output throughput on TP8.

There are three reasoning modes: Low, Medium and High. High is the default.

If you want a refresher on where a model like this sits, I wrote about [what a frontier model actually is](https://www.oguzhan.co/what-is-a-frontier-model/).

## What does Mistral Large 4 cost?

This is the part I would look at first. Mistral lists these API prices:

| Token type | Price per 1M tokens |
| --- | --- |
| Input | $0.70 |
| Cached input | $0.07 |
| Output | $2.10 |

Cached input at a tenth of the normal input price matters a lot for agent loops and long system prompts, where the same context gets sent again and again. If you compare cheap models often, my post on [MiniMax M3.1 Flash pricing](https://www.oguzhan.co/minimax-m3-1-flash-space-bunny-price/) uses the same per-token lens.

## Benchmarks, read with care

All of these are Mistral’s numbers from the announcement. No technical report yet. Mistral says it is “coming soon”.

| Benchmark (Mistral’s figures) | Mistral Large 4 |
| --- | --- |
| AA Coding Agent Index | 48% (GLM-5.3 tie, 47% fewer tokens) |
| Terminal-Bench 4.0 | 41% |
| FrontierCode 2.0 | 19.7% |
| AA Cyber Index | 51% |
| BrowseComp | 84.5% |

A few notes from the same page:

- On the Artificial Analysis Coding Agent Index, Mistral says it is “effectively tied” with GLM-5.3 while using “47% fewer tokens per task on average”. Fewer tokens per task plus a low output price is the combination worth watching.
- Terminal-Bench 4.0 went up 13 points over the preview, with 30% fewer output tokens.
- Cognition’s FrontierCode 2.0 is a private benchmark, and Mistral says Large 4 leads it at 19.7%.
- On Finch (FinWorkBench) it scores 73.3%, second to Kimi K3 among open models. B3 Agent Security rose to 97.1% from 93.3% in the preview.
- Average refusal on malicious cyber requests is 96.6%.
- In Surge’s blind code preference test, Mistral says Large 4 beat GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, and only Claude Opus 5.5 came out ahead.

That is a strong set of claims. I would wait for independent runs before treating any single number as settled.

## How to run it: API, managed, self-host

**1. API, the easy way.** Use the Mistral API, Mistral Studio or OpenRouter, and pick a reasoning mode. Mistral also lists it in Vibe Code. For most projects this is where I would start.

**2. Managed.** Microsoft Foundry and NVIDIA NIM both carry it. Hugging Face has an Inference Endpoints recipe on 8x H200. The [Hugging Face blog](https://huggingface.co/blog/mistral-large4) reports about 0.6 s time to first token for a 1K prompt, around 65 tokens per second for a single request, and 120 to 138 tokens per second with DSpark on chat.

**3. Self-host.** The weights are on Hugging Face and you serve them with vLLM or SGLang. As of the vLLM recipe there is no release wheel yet, so you need the `vllm/vllm-openai:mistral-large-4` Docker image. The FP8 command on 8 GPUs from the vLLM blog:

`vllm serve mistralai/Mistral-Large-4-1T-A52B --tensor-parallel-size 8 --speculative-config '{"method": "dspark", "num_speculative_tokens": 3}' --tool-call-parser mistral --enable-auto-tool-choice --reasoning-parser mistral`

The [SGLang cookbook](https://docs.sglang.io/cookbook/autoregressive/Mistral/Mistral-Large-4) has verified FP8 at TP8 on 8x B300 and NVFP4 at TP4 on 4x GB300. It recommends temperature 1.0 and top_p 0.95.

Context is advertised at 1,048,576 tokens. vLLM has validated it to 256k, so I would plan around 256k until more testing lands.

Unsloth and the llama.cpp team have published GGUFs. I have not tested them and the sources do not give hardware numbers for them, so I am not going to guess. If you like local setups, my [DeepSeek DeepInfer post](https://www.oguzhan.co/deepseek-deepinfer-eagle-4-muse-glimmer/) covers that side of things.

![Isometric illustration of a chunky orange cube being unloaded from a crate onto a GPU server](https://oguzhanco.s3.eu-west-3.amazonaws.com/wp-content/uploads/2026/10/30082846/mistral-large-4-le-chonk-gpu-server.webp)

*Le Chonk needs a full eight-GPU node, not a workstation.*

## Hardware: what you actually need

Here is the honest bill. There is no validated single consumer GPU route.

| Checkpoint | Size | Hardware validated |
| --- | --- | --- |
| FP8 | ~1.06 TB | 8x B300, 2×8 H200, GB300 |
| NVFP4 | ~610 GB | Blackwell: 8x B200, GB300 |
| MXFP4 (AMD Quark) | ~578 GB | 4x MI350 series |

The vLLM recipe says FP8 needs about 128 GiB of weights per GPU across 8 GPUs. NVFP4 drops that to about 76 GiB, but it only runs on Blackwell. The AMD Quark MXFP4 build fits on 4x MI350 series cards. The checkpoint names are `mistralai/Mistral-Large-4-1T-A52B` for FP8 and `mistralai/Mistral-Large-4-1T-A52B-NVFP4` for NVFP4.

If you do not already have nodes like these, the per-token price above will almost certainly beat renting them.

![Grid of 128 glowing nodes with four lit in orange, showing mixture-of-experts routing](https://oguzhanco.s3.eu-west-3.amazonaws.com/wp-content/uploads/2026/10/30082848/mistral-large-4-moe-routing-experts.webp)

*128 routed experts, top-4 per token plus one shared expert.*

## The license catch

Mistral uses the Mistral Ecosystem Open-Weight license, which it describes as “a modified MIT license”. You can use, modify, fine-tune and commercialize the model for free. The one condition: if you run it as a Model-as-a-Service provider, you need a separate commercial license from Mistral.

So for a company building its own product on top of it, this is open in practice. For a hosting company that wants to resell API access, it is not. Weights and how they get shared is a topic I have followed before. My post on [China’s export controls for model weights](https://www.oguzhan.co/china-ai-export-controls-model-weights/) and the [AI weekly for 19 to 25 October](https://www.oguzhan.co/ai-weekly-19-25-oct-2026-fences-and-proofs/) give the background.

## FAQ

**Is Mistral Large 4 open source?**\
 It is open weight under a modified MIT license. Model-as-a-Service providers need a separate commercial license.

**Can I run Mistral Large 4 on one GPU?**\
 No validated single-GPU setup exists. FP8 needs about 128 GiB per GPU across 8 GPUs.

**How many active parameters does it have?**\
 49B per token, or 52B including embeddings and output.

**What is the context window?**\
 1M tokens advertised. vLLM has validated 256k.

**How much does it cost?**\
 $0.70 per million input tokens and $2.10 per million output tokens. Cached input is $0.07.

**How many languages does it support?**\
 Mistral says 160+.

## Sources

- [Mistral, Mistral Large 4 open weights](https://mistral.ai/news/mistral-large-4-oss/)
- [vLLM blog, Mistral Large 4](https://vllm.ai/blog/2026-10-29-mistral-large-4)
- [vLLM recipe, Mistral-Large-4-1T-A52B](https://recipes.vllm.ai/mistralai/Mistral-Large-4-1T-A52B)
- [Hugging Face blog, Mistral Large 4](https://huggingface.co/blog/mistral-large4)
- [SGLang cookbook, Mistral Large 4](https://docs.sglang.io/cookbook/autoregressive/Mistral/Mistral-Large-4)