Mistral Large 4

Learning Objectives

• Understand what Mistral Large 4 is and how it differs from Mistral Large 3
• Read the Mistral Ecosystem Open-Weight License correctly, including the one group of users it restricts
• Judge where Large 4 leads, where it trails, and which of its results are Mistral's own

What Is Mistral Large 4?

Mistral Large 4 is the flagship model from Mistral AI, the Paris-based lab. It is a natively multimodal mixture-of-experts model with 1 trillion total parameters, of which 49 billion are active for each token (52 billion counting the embedding and output layers). It supports a 1 million-token context window, roughly four times that of Mistral Large 3, and Mistral says it was trained from scratch on 3,800 Nvidia Grace Blackwell chips in Mistral's own European data centers.

Mistral opened a public preview through its API on October 6, 2026 and released the weights on October 29, 2026, after about three more weeks of reinforcement learning. The released model is not identical to the preview: Mistral reports large gains in coding and in defensive security work since then, and a new set of three reasoning modes (low, medium and high, with high as the default) for trading cost against depth. Mistral's informal nickname for it is "Le Chonk."

It succeeds Mistral Large 3, a 675 billion-parameter model with 41 billion active parameters, which remains available and is now the previous generation.

The License: Free for Most, Not for API Resellers

Large 4 ships under a new license, the Mistral Ecosystem Open-Weight License. It is written as a modified MIT license, and for most readers it behaves like one: you may use, copy, modify, fine-tune and sell products built on the model, with the usual requirement to keep the copyright notice.

The modification is one clause. Any company that runs a "Model as a Service" business, meaning it gives third parties access to the model for inference or further training in a way that lets them control the inputs, parameters or training data, must sign a separate commercial license with Mistral before using it commercially. Two things are carved out of that definition: end-user products where the model sits inside a specific feature, and simply relaying requests to a model someone else hosts. Internal use, where nothing is made available to outsiders, is exempt, as is any use through Mistral's own products.

Unlike several recent open-weight licenses, it sets no revenue threshold and excludes no country. The practical upshot: a company building its own product or running the model in-house can treat it as open, while an inference provider or cloud reselling it as an API needs to talk to Mistral first. That is why this catalog lists it as freemium rather than open source.

Pricing & Access

On Mistral's own API, Large 4 costs 70 cents per million input tokens and two dollars ten per million output tokens, with cached input at seven cents per million. Because reasoning tokens count toward output, the reasoning mode you choose drives the real bill: Mistral recommends the low mode for simple tasks and keeps high as the default for the hardest ones.

What Mistral Says It Does Well

Mistral leads with two uses, coding agents and defensive cybersecurity. All of the figures below are Mistral's own reported results and had not been independently reproduced at release.

• Coding: Mistral reports the best open-weight score on Cognition's private FrontierCode 2.0 benchmark, 19.7 percent, and 41 percent on Terminal-Bench 4.0, 13 points above its own preview. On the Artificial Analysis Coding Agent Index it scores 48 percent, effectively tied with GLM-5.3, while using about 47 percent fewer tokens per task on average.
• Defensive security: It scores 51 percent on the Artificial Analysis Cyber Index, which tests finding, reproducing and patching vulnerabilities, and Mistral says it does this work without needless refusals. Mistral also reports a 96.6 percent average refusal rate on malicious cyber requests, the highest it measured among open models.
• Other tasks: Mistral reports 84.5 percent on the BrowseComp web-research test, and 73.3 percent on the Finch finance benchmark, which it says is second among open models to Kimi K3.

Where It Still Trails

Mistral's own charts show where Large 4 is not the leader, and they matter more than the headline claims.

• On security, it is not the top open model. In Mistral's Cyber Index chart, two other open-weight models, DeepSeek V4.1 Flash and Xiaomi's MiMo-V2.6-Pro, score higher than its 51 percent. Its edge is the combination of solid security skills with high refusal of malicious requests, not raw capability.
• On coding, closed models still lead. In Surge's blind human-preference test of code answers, graders preferred Large 4 over every open-weight rival tested, but preferred Claude Opus 5.5's answers about 78 percent of the time.
• It is very large to run yourself. Only 49 billion parameters are active per token, but self-hosting means loading the full trillion into memory, which takes multi-GPU servers well beyond most teams' hardware.
• Results are vendor-reported. Benchmarks, refusal rates and preference tests all come from Mistral's launch materials.

Mistral Large 4 vs Mistral Large 3

Large 3 stays a reasonable choice where its smaller size matters, or where an organization has already validated it.

Best Use Cases

Getting Started

1. Try it hosted first — call it through the Mistral API or Mistral Studio before committing to hardware
2. Pick a reasoning mode — start with medium for everyday tasks and reserve high for the hardest problems
3. Check the license — confirm whether your use counts as a Model-as-a-Service business before hosting it for others
4. Self-host if needed — download the weights from Hugging Face and serve them with vLLM or SGLang on multi-GPU hardware
5. Benchmark on your own work — the launch figures are Mistral's, so compare it with your current model on your own tasks

Key Takeaways

• Mistral Large 4 is Mistral's flagship open-weight model: 1 trillion parameters, 49 billion active, a 1 million-token context window and native multimodal input
• Mistral previewed it on October 6, 2026 and released the weights on October 29, 2026, as it had promised
• Its license is free for most uses, but companies that resell API or fine-tuning access must sign a separate commercial license, with no revenue threshold and no excluded countries
• Mistral reports leading open-weight coding results and strong defensive security skills, but two open rivals score higher on its own security chart and graders still preferred Claude Opus 5.5 for code
• It succeeds Mistral Large 3, which remains available as the smaller previous generation