EN

MiniMax: a Chinese multimodal AI maker with open models

MiniMax is a Chinese company building AI models for text, voice, video and music, along with consumer apps. Its MiniMax M family of language models is released as open weights, with a strong focus on coding and agents.

Data as of October 22, 2026

Key figures

CountryChina
Founded2022
TypeStart-up
Weights policyOpen weights
Model familiesMiniMax M
Consumer assistantMiniMax Agent
Official siteminimax.io · Developer API
Models tracked here6

Built multimodal from day one

MiniMax describes itself as a global foundation-model company, founded in early 2022, with the motto “intelligence with everyone.” Unlike labs that started with text alone, it built models from the outset that handle several kinds of content: text, audio, images, video and music. That is what multimodal AI means.

You can see this in its products. Consumers know MiniMax for Talkie and Xingye (character chat apps), for the Hailuo video generator and for its synthetic voice models, offered to developers through the API.

The MiniMax M family, built for coding and agents

On the large language model side, the MiniMax M family is designed for software development and AI agents. The M2 and then M3 generations focus on efficiency: they are mixture-of-experts models that activate only part of their parameters for each token, which cuts compute costs. The latest version is multimodal, and the “Flash” variant targets high-volume workloads.

Open weights, with an attribution clause

MiniMax publishes its M language models as open weights on Hugging Face. The license is based on MIT with one extra condition: very large commercial products must display the model name in their interface. The company also keeps proprietary versions, notably for voice, and offers its models through its API.

In practice, this openness lets businesses test MiniMax M on their own hardware, or use a third-party host, without being locked into a single provider. Keep in mind that running the largest M models yourself requires substantial server hardware; for a first test, the API or a hosted version is usually far simpler. As with any Chinese provider, also check where your data is processed if you have compliance obligations.

How to use MiniMax

  • MiniMax Agent and MiniMax Code, for direct use in your browser.
  • The API platform (platform.minimax.io), for text, voice, video and music.
  • Downloading the open models for self-hosting.

To compare MiniMax M with other open models, see our open-source models page.

MiniMax models tracked on this site

ModelsReleasedCapability indexPrice (input / output)
MiniMax-M3.1-FlashOctober 19, 2026—$0.15 / $0.60
MiniMax-M3June 1, 202655.5$0.30 / $1.2
MiniMax-M2.7March 18, 2026—$0.30 / $1.2
MiniMax-M2.5February 12, 2026—$0.30 / $1.2
MiniMax-M2.1December 23, 2025—$0.30 / $1.2
MiniMax-M2October 27, 2025—$0.30 / $1.2

Frequently asked questions

When was MiniMax founded?

According to its official website, MiniMax was founded in early 2022. Some third-party sources say late 2021; the gap likely reflects the difference between legal incorporation and the start of operations.

Can I use MiniMax M models freely?

They are published on Hugging Face under an MIT-style license with one added condition: above a very high commercial usage threshold, the product must display the model name. For most projects that means very permissive use, but read the license file.

What does MiniMax offer besides language models?

MiniMax also builds text-to-speech models (Speech), video generation (H3, known through Hailuo) and music generation (Music). It runs apps such as Talkie, Xingye, MiniMax Code and MiniMax Audio.

How do I access the models through an API?

Through MiniMax's open platform (platform.minimax.io internationally). The open models are also offered by third-party hosts and can be installed on your own servers.

Data sources

Data as of October 22, 2026

Report an error

Does a figure, a date or a fact look wrong? Tell us: every report is checked.