Mistral Large 4: Specs, Benchmarks, API Price & License

Mistral Large 4 is the largest open-weight model ever shipped by a European lab, and the first Mistral model built to compete with the top open models from China. Mistral announced Mistral Large 4 on October 6, 2026 as a public preview API, then released the weights on October 29, which is when search interest jumped.

This guide covers what it is, its specs, benchmarks, API price, license and how to run it, with numbers from Mistral's launch post, the official model docs and the Hugging Face release.

Mistral Large 4 at a glance

What is Mistral Large 4?

Mistral Large 4 is a general-purpose model with reasoning always on. It reads text and images, writes text, calls functions, returns structured output and runs agents through Mistral's Agents API. Mistral calls it "our largest and most capable model to date" and says it will be the base for a new family of specialized Mistral models.

Two things make Mistral Large 4 stand out. First, it is open weight at frontier scale, so companies can run it on their own servers. Second, Mistral trained it for security work that closed models often refuse, like reproducing a real bug before patching it. More than 160 languages were in the training data, including every official EU language.

Mistral Large 4 specs and architecture

Hugging Face's launch post gives the full layout:

• 72 decoder layers, hidden size 6,144, 64 attention heads.
• Mixture of experts: 128 experts, a router picks 4 per token, plus 1 shared expert, so 5 experts run per token. The first layer is dense.
• Two kinds of sparsity: Multi-head Latent Attention with a DeepSeek Sparse Attention indexer that picks up to 2,048 earlier positions for each query.
• Vision: a 1.6B Pixtral encoder (30 layers, 14×14 patches).
• Reasoning effort: low, medium or high. High is the default and reasoning cannot be turned off.

Mistral Large 4 benchmarks

All scores below are from Mistral's launch post and the Hugging Face blog. Several are independent tests run by Artificial Analysis and Vals AI.

In a blind coding review with Surge AI, Mistral Large 4 Preview scored 3.74 out of 5, second of five models, behind Claude Opus 5 (4.22) and ahead of Kimi K3 (3.59) and GLM-5.3 (3.60). As of October 29, Hugging Face says it was the fifth-best open-weight model on the Artificial Analysis Intelligence Index. Vals AI ranks the open-weight build #17 of 30 on its Vals Index at 56.42%.

Mistral Large 4 pricing and API

On Mistral Studio, Mistral Large 4 costs $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. The API model ID is mistral-large-4. Features listed in the docs: chat completions, function calling, structured outputs, document Q&A, batching, prefix, and the Agents and Conversations API with built-in tools. You can also call it through NVIDIA NIM and Hugging Face Inference Providers.

For context, that puts Mistral Large 4 well below closed frontier models on price and in the same range as big Chinese open models. Mistral also runs a European deployment it operates end to end under EU law, which matters to banks, governments and other buyers who care where data goes.

Mistral Large 4 license

Mistral Large 4 ships under the Mistral Ecosystem Open-Weight License. Hugging Face describes it as an MIT-derived license that adds limits on model-as-a-service offerings. In plain words: you can download the model, fine-tune it and run it inside your own products, but reselling it as a hosted API has extra conditions. Read the license on the model card before you build a hosting business on it.

How to run Mistral Large 4

• Fastest: call mistral-large-4 on Mistral Studio or try it in the Mistral playground.
• vLLM: the NVFP4 checkpoint mistralai/Mistral-Large-4-1T-A52B-NVFP4 runs on 8 NVIDIA Blackwell GPUs with tensor and expert parallelism, plus a DSpark drafter for speed.
• Transformers: use AutoModelForMultimodalLM and launch with torchrun --nproc-per-node 8.
• llama.cpp: GGUF builds from Unsloth and ggml-org, for example llama serve -hf ggml-org/Mistral-Large-4-1T-A52B-GGUF:Q2_K.
• Dell Enterprise Hub: a ready deploy on 4× B300 GPUs.

Self-hosting is a data-center job. Even the 4-bit build needs a full multi-GPU node, so most teams will start with the API and move to their own hardware later. The weights and quants are in the Mistral Large 4 Hugging Face release.

Mistral Large 4 vs other open models

Pick Mistral Large 4 if you need security work without refusals, strong visual grounding, or an EU-hosted open model. Look at DeepSeek or Kimi if raw coding scores per dollar are all that matters.

Put Mistral Large 4 to work with a Sumus AI Employee

Yes, you can test Mistral Large 4 against the models you already use without writing a harness. This post was written by a Sumus AI Employee: it read Mistral's launch post, the docs and the Hugging Face release, pulled the numbers and built the chart above. Give a Sumus AI Employee a task like "run our 20 support prompts on Mistral Large 4 and Claude and show where each wins", and it plans the work, brings in other agents and hands back results your team reviews in the same chat.

What we're watching

• Specialized models built on Mistral Large 4 that Mistral says are coming.
• Gains from the RL run, which Mistral says is still in flight.
• Independent scores on Artificial Analysis and LMArena for the open-weight build.