Mistral Large 4 Guide: Open Weights, API, Pricing, Setup

Last updated: October 29, 2026

Mistral Large 4 (“Le Chonk”) went open weight on October 29, 2026, with an improved model that leads open-weight models on Terminal-Bench 4.0 and FrontierCode. It is a 1.05T-parameter multimodal mixture-of-experts model with 52B active parameters, a 1M-token context window and API pricing of $0.70 input and $2.10 output per million tokens.

Mistral first released the model as a public preview on October 6 (preview announcement). The open-weight launch adds reasoning modes, better token efficiency and downloadable checkpoints (open-weight announcement). For earlier Mistral news, see our explainer on the physics model CERN and Mistral plan to build.

Contents

What’s new in the open-weight release
Specs and API features
How to use Mistral Large 4
Running the open weights
Pricing and license
How it compares
Limitations to know
Frequently asked questions

What’s new in the Mistral Large 4 open-weight release

Mistral says three weeks of additional reinforcement learning after the preview produced large gains in coding and defensive cybersecurity, along with better reasoning efficiency that lowers inference costs. The table compares figures Mistral published for the preview and for the October 29 release. These are vendor-reported results, not independent tests.

Item	Preview (Oct 6)	Open-weight release (Oct 29)
Terminal-Bench 4.0	28.3%	41%, with 30% fewer output tokens
B3 Agent Security (attack resistance)	93.3%	97.1%
Reasoning modes	Reasoning plus a None mode	Low, Medium, High (default)
Weights	Not yet released	FP8 and NVFP4 on Hugging Face

Other figures in the release post: 19.7% on Cognition’s private FrontierCode 2.0, which Mistral says is the top score; 48% on the Artificial Analysis Coding Agent Index, about tied with GLM-5.3 while using 47% fewer tokens per task; 51% on the Artificial Analysis Cyber Index; 84.5% on BrowseComp; and 73.3% on the FinWorkBench (Finch) spreadsheet benchmark. Mistral also reports a 96.6% average refusal rate on malicious cyber prompts, the highest it measured among open models.

Specs and API features

These details come from the Mistral Large 4 model page and the Hugging Face launch post:

Item	Mistral Large 4
API model name	`mistral-large-4`
Parameters	1.05T total, 52B active, plus a 1.6B Pixtral vision encoder
Inputs and output	Text and images in, text out
Context window	1M tokens
Architecture	72 layers; 128 experts with 4 routed plus 1 shared per token; sparse attention selecting up to 2,048 key positions
API features	Structured outputs, function calling, document Q&A, prefix, batching, agents and conversations, built-in tools
Languages	More than 160, including every official EU language
License	Mistral Ecosystem Open-Weight License

Mistral trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers, according to the preview post.

How to use Mistral Large 4

The model is available in Mistral Studio, the Mistral API, OpenRouter, Mistral’s Vibe Code agent, NVIDIA-hosted endpoints and Microsoft Foundry. To get started with the hosted API:

Create an API key in Mistral Studio.
Set model to mistral-large-4 in a /v1/chat/completions or /v1/conversations request.
Pick a reasoning mode. Use Low for simple tasks, Medium for everyday work and High (the default) for hard problems that need precision.
Send images alongside text for document, chart or diagram questions.
Use /v1/batch for large offline jobs, and structured outputs when you need JSON that matches a schema.
Try it in a repository with Vibe Code by giving it a feature to build or a bug to fix.

Hugging Face shows the reasoning control as a reasoning_effort field on its OpenAI-compatible Inference Endpoints; the answer is returned in choices[0].message.content and the reasoning in choices[0].message.reasoning. If you build coding agents, our Claude Code power-user guide covers a comparable agent workflow.

Running the Mistral Large 4 open weights

The checkpoint is mistralai/Mistral-Large-4-1T-A52B on the Hugging Face model page, with an NVFP4 variant. Hugging Face lists day-one support in Transformers, vLLM and llama.cpp, plus fine-tuning with TRL. The vLLM recipe for the NVFP4 checkpoint, which requires NVIDIA Blackwell GPUs, starts like this:

vllm serve mistralai/Mistral-Large-4-1T-A52B-NVFP4 \
  --config-format mistral --load-format mistral \
  --tensor-parallel-size 8 --enable-expert-parallel \
  --max-model-len 131072 --enable-prefix-caching

Practical notes from the Hugging Face post:

Speculative decoding: the model ships with a DSpark draft head. On an 8 x H200 Inference Endpoint, Hugging Face measured about 65 tokens per second for a single request, rising to 120–138 on chat prompts where drafts are accepted.
GGUF quants: the llama.cpp team publishes IQ2_XXS, Q2_K, Q4_K_M, NVFP4, Q8_0 and BF16. The 2-bit quants run on one Apple Silicon Ultra with 512 GB of unified memory; Q2_K on an M3 Ultra generated about 15 tokens per second.
Reasoning: it cannot be disabled; none is treated as high.
Customization: Mistral Forge supports enterprise tuning, and TRL supports LoRA fine-tuning on the FP8 weights.

Pricing and license

Mistral lists these API prices on the model page:

Per 1M tokens	Price (USD)
Input	$0.70
Cached input	$0.07
Output	$2.10

The weights use the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT license. You can use, modify, fine-tune and commercialize the model, but a business that operates a Model-as-a-Service offering must get a separate commercial license before using it commercially. Read the license on the model page before you host it for third parties.

How Mistral Large 4 compares

Mistral says the model leads open-weight models in coding and defensive security, and that on the Artificial Analysis Cyber Index it leads open-weight models developed outside China by a wide margin. In Surge Code blind pairwise tests reported by Mistral, graders preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking it behind only Claude Opus 5.5. For the closest open-weight rival, read our DeepSeek V4.1 Flash guide. For closed alternatives, see our GPT-6.1 Terra guide and Gemini 4 Argon guide. Another recent open release is covered in our Muse Spark Open explainer.

Limitations to know

Most benchmark figures come from Mistral and its partners; check them on your own workloads.
The hardware bar is high: full-precision serving needs a multi-GPU node, and long contexts slow local generation sharply.
Mistral says it is still adjusting cyber guardrails after seeing potential uplift for real attacks in preview traffic.
A version with reduced moderation for security research is limited to cybersecurity leaders, vetted partners and state authorities during red-teaming.

Frequently asked questions

What is Mistral Large 4?

Mistral Large 4, nicknamed Le Chonk, is Mistral AI’s largest model. It is a natively multimodal mixture-of-experts model with 1.05T total parameters, 52B active parameters, a 1.6B vision encoder and a 1M-token context window. It entered public preview on October 6, 2026.

When were the Mistral Large 4 weights released?

Mistral released the open weights on October 29, 2026, together with an improved version of the model trained with about three more weeks of reinforcement learning. The weights are on Hugging Face in FP8 and NVFP4 variants.

How much does the Mistral Large 4 API cost?

Mistral lists $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens for mistral-large-4.

Can I use Mistral Large 4 commercially?

Yes, under the Mistral Ecosystem Open-Weight License, a modified MIT license. Mistral says the model is free to use, modify, fine-tune and commercialize, but businesses that run a Model-as-a-Service offering need a separate commercial license from Mistral.

Can I turn off reasoning in Mistral Large 4?

No. The open-weight release offers low, medium and high (default) reasoning modes, and Mistral says Low replaces the old None mode. Hugging Face notes that a reasoning_effort value of none is treated as high in Transformers.

What hardware do I need to run Mistral Large 4 locally?

It is a very large model. Hugging Face reports that the 2-bit GGUF quants run on a single Apple Silicon Ultra machine with 512 GB of unified memory, while other formats need multi-GPU or distributed setups, such as eight NVIDIA H200 or Blackwell GPUs.

Sources

Image: Mistral AI logo, public domain, via Wikimedia Commons. Placed on a white background and resized.

Mistral Large 4: Power-User Guide to Open Weights, API, Pricing and Setup