AI Engineer
← All organizations

Multimodal AI models and applications

MiniMax

MiniMax develops AI foundation models for coding and generating text, audio, images, video and music. Its applications include MiniMax Code, MiniMax Design, MiniMax Audio and Talkie, while its Open API Platform gives enterprises and developers access to its models. Its generation capabilities include video with native stereo audio and music composed from a creative concept and optional lyrics, connecting model research to software development and creative work.

Founded in early 2022, MiniMax’s founders include Yan Junjie, its current CEO and CTO, and Yeyi Yun, its current COO. Its research includes MiniMax-01, which combines Lightning Attention with a mixture-of-experts architecture to support long contexts. The 2025 technical report describes training with one million tokens of context and extending to four million at inference, supported by optimized distributed computation.

MiniMax completed its Hong Kong listing in January 2026, raising approximately HK$4.8 billion in gross IPO proceeds. It earns revenue from AI applications and platform and enterprise services, including paid API usage and Token Plans. As of August 2026, the company reported cumulatively serving more than 300 million individual users and more than two million enterprises and developers, illustrating the reach of both its applications and model platform.

www.minimax.io ↗

2 talks

Newest first

1 speaker at AIE

Affiliations reflect their AIE appearances, not necessarily current employment.

Olive SongRL Lead

Start here

  1. Minimax M2

    Start here to learn how MiniMax-M2 interleaves reasoning and tool use, and how Olive Song presents economical multi-agent workflows.

    Olive SongAI Engineer Code 2025

  2. Agents at Scale: Inside MiniMax's Model and the Infrastructure Behind It

    Turn to this discussion for the partnership and infrastructure considerations involved in serving an open-weight multimodal model at scale.

    Dan Fu · Olive SongAI Engineer World's Fair 2026

Messages from the stage

Training agents for varied working conditions

The MiniMax-M2 presentation examines developer-informed reinforcement learning and expert reward models, alongside robustness across prompts and environments.

Evaluating and serving extended agent tasks

The joint MiniMax-M3 discussion connects benchmark selection, including KernelBench and OSWorld, with sparse attention and KV-cache challenges in GPU inference.

Affiliations reflect each recorded session, not necessarily current employment. The MiniMax-M3 session is a joint discussion with Together AI's Dan Fu.

Company sources · checked 2026-08-28