dalao.ai

Mistral AI

Frontier open-weight models you can run on your own infrastructureReference profile

About Mistral AI

Mistral AI is a French company, registered with the Paris Trade and Companies Register and founded in April 2023. It publishes open-weight models and also sells a paid API and Vibe, an agent product formed when its Le Chat assistant and its Mistral Code coding product were merged and renamed. On 2026-10-29 it released the weights of its 1-trillion-parameter flagship, Mistral Large 4, under a license that is not on the OSI's list of approved licenses.

EditorialUpdated 4 min read

What problem it solves

Mistral positions itself around "sovereign AI". The Large 4 launch post defines sovereignty as organizations keeping control of the models they depend on and the intelligence they create: they can inspect and customize the model and deploy it on their own infrastructure, without depending on Mistral to serve it through an API. Commercially, Mistral earns from the API, enterprise deployments and subscriptions.

Its docs group the products into three parts: Vibe (a unified agent for everyday work and coding), Studio (the developer console and API) and Admin (organization, billing and access management). The App Store listing is "Vibe by Mistral (ex-Le Chat)" and its release notes say "Le Chat is now Vibe"; the company timeline marks Mistral Code, launched 2025-06-04, as "later Vibe".

Mistral Large 4 ranks as the best open-weight model for coding and among the world's most capable open models overall.

Mistral blog, 2026-10-29 (the company's own claim)

Core capabilities

  • Mistral Large 4: the model card describes a natively multimodal mixture-of-experts model with 1 trillion total parameters and 49B active per token (52B counting embedding and output layers), trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters and fluent in 160+ languages.
  • Reasoning modes: Large 4 offers Low, Medium and High; High is the default.
  • Other models: the docs home page also lists Mistral Small 4 (Apache-2.0) and third-party models hosted on Mistral's platform, such as GLM-5.3 and DeepSeek V4.1 Flash.
  • Vibe: on the web, iOS and Android, in the terminal, and as VS Code and JetBrains plugins, with a coding agent, remote sandboxes, scheduled tasks, spreadsheet and slide generation, connectors and custom MCP connectors (beta).

Who uses it

Mistral's site and docs give no user numbers for Vibe or the API.

The Large 4 launch post quotes people from 20 organizations, all statements Mistral chose to publish. The "Built for the Defenders" (cybersecurity) section has 11: Booz Allen Hamilton, Cisco, Cloudflare, CrowdStrike, Elastic, ENISA (the EU Agency for Cybersecurity), Fortinet, NVIDIA, Orange Cyberdefense, Rubrik and Zscaler. "Built for Work" has one, La Banque Postale. "Built for Sovereignty" has 8: ABN AMRO, ASML (the text gives only the speaker's title; the company is shown by its logo), Belfius, Crédit Agricole, the European Space Agency (ESA) Φ-lab, Hugging Face, ZML and Microsoft. As of 2026-10-30 the main Large 4 repository on Hugging Face has about 1,233 likes.

How it compares

The charts in Mistral's launch post compare Large 4 with GLM-5.3, GLM-5.3 Flash, Kimi K3, DeepSeek V4.1 Flash, DeepSeek-V4-Pro, MiMo-V2.6-Pro, Gemma-4-31B, Beam 1.0 and Muse Spark Open, and the human-preference test also includes Claude Opus 5.5. On the Artificial Analysis Coding Agent Index it cites, Large 4 scores 48%, roughly level with GLM-5.3, while using 47% fewer tokens per task on average. Mistral chose and cited these figures.

The licenses differ too. Mistral's docs list GLM-5.3 and DeepSeek V4.1 Flash as MIT-licensed, while Large 4 restricts commercial Model-as-a-Service use (see Pricing & license).

Who it's for / Not for

Good fit:

  • European companies and public bodies that care about data sovereignty and need a large model running in their own facilities;
  • teams that want to use and fine-tune a large model internally, which the commercial-license clause doesn't cover.

Poor fit:

  • inference providers who want to resell Large 4 as an API: they must sign a commercial license with Mistral first;
  • individual developers with a few GPUs, for whom a 1-trillion-parameter model is out of reach.

Getting started

Individuals can sign up at mistral.ai and try Vibe for free. Developers get an API key in Studio; Large 4 is called mistral-large-4 in the API.

To self-host, download the weights from Hugging Face. The model card lists support for vLLM, SGLang, transformers and llama.cpp.

Pricing & license

As of 2026-10-30:

  • Vibe: a Free tier; Pro at $14.99 a month before tax, or $5.99 a month on the student plan; Team at $24.99 per user per month before tax; Enterprise through sales.
  • API: Large 4 costs $0.70 per million input tokens, $0.07 for cached input and $2.10 for output. Per the pricing FAQ, batch processing halves the price and cached input cuts input cost by up to 90%.

Licenses differ by model:

  • Mistral Large 4: the Mistral Ecosystem Open-Weight License, which Mistral calls a modified MIT license. It allows use, modification, redistribution and commercial use, but a company running a Model-as-a-Service business needs a separate commercial license before using it commercially. Internal use is exempt.
  • Mistral Small 4: Apache-2.0.
  • Third-party models hosted on Mistral's platform, GLM-5.3 and DeepSeek V4.1 Flash: listed as MIT in the docs.
  • As of 2026-10-30, of the 83 models under mistralai on Hugging Face, 66 are tagged apache-2.0, 15 "other", 1 cc-by-nc-4.0 and 1 has no license tag.

Risks & limits

How open "open" is. Mistral markets Large 4 as fully open-weight, but its license is not on the OSI's list of approved licenses and it restricts commercial Model-as-a-Service use, so "open-weight" is the accurate term.

Pricing page wording. The FAQ's example price (Mistral Large at $0.5 input and $1.5 output per million tokens) doesn't name a model version and differs from the Large 4 launch price of $0.70 and $2.10. The FAQ also says Apache 2.0 models such as Mistral 7B need a Mistral license for commercial deployment, which doesn't match the Apache-2.0 terms, which allow commercial use. The LICENSE shipped with each model is what counts.

Self-chosen benchmarks. The comparison charts in the launch post use opponents and metrics Mistral picked.

Our verdict

Large 4 takes Mistral's open-weight models to the trillion-parameter scale, with "sovereign AI" as the main pitch. Keep the licenses apart, though: most of the smaller models are Apache-2.0, while the flagship Large 4 uses a custom license with commercial restrictions. Anyone in the inference-hosting business should read the LICENSE first.

Why it matters

Mistral AI ships both open-weight and commercial models and runs the Le Chat assistant.

Sources

Official

  1. Mistral AI: Aboutmistral.ai · Accessed Oct 30, 2026
  2. Mistral blog: Mistral Large 4 open-weight releasemistral.ai · Accessed Oct 30, 2026
  3. Mistral documentationdocs.mistral.ai · Accessed Oct 30, 2026
  4. Mistral pricingmistral.ai · Accessed Oct 30, 2026
  5. Mistral AI: Legal noticemistral.ai · Accessed Oct 30, 2026

Media & third party

  1. mistralai/Mistral-Large-4-1T-A52B model cardHugging Face · Accessed Oct 30, 2026
  2. Hugging Face: models by mistralaiHugging Face · Accessed Oct 30, 2026
  3. Mistral Large 4 LICENSE (Mistral Ecosystem Open-Weight License)Hugging Face · Accessed Oct 30, 2026

Updates & reading

NewsMistral

Mistral releases open weights for Mistral Large 4, its 1-trillion-parameter flagship, on Hugging Face

On Oct 29 Mistral released the weights of Mistral Large 4 ("Le Chonk"), in API preview since early October. The multimodal MoE model has about 1T total and 49B active parameters and three reasoning modes. Mistral says three more weeks of RL lifted Terminal-Bench 4.0 from 28% to 41%, calling it the best open-weight coding model (vendor claim). A quantized version runs on four Nvidia B200/B300 GPUs.

3 sources · 2 posts on X