MiniMax

MiniMax is an AI company that builds foundation models for text, video, speech and music and sells access to them through its own apps and an API platform. It describes itself as a “global AI foundation model company” founded in early 2022, and its site lists the language models MiniMax M3, M2.7 and M2.5, the video model MiniMax H3, MiniMax Speech 2.8 and MiniMax Music 3.0. MiniMax says its models and products have served more than 300 million individual users and more than 2 million enterprises and developers Vendor-stated.¹ The ATOM Report, a study of open language models, counts it among China’s open-model developers.²

Its open-weight releases on Hugging Face began with MiniMax-Text-01 and MiniMax-VL-01 in January 2025,³ and Hugging Face’s spring 2026 report names MiniMax among the organizations that had favored closed approaches and then “shifted decisively toward open releases”.⁴ Its current flagship language model, MiniMax-M3, was announced on June 1, 2026: a natively multimodal model with about 428 billion parameters, about 23 billion of them activated, and a 1M-token context Vendor-stated.⁵⁶ Licenses differ by model. MiniMax-M2.7 ships under a non-commercial license, so any commercial use needs MiniMax’s prior written authorization; M3’s MiniMax Community License allows commercial use, but products or services with more than $20 million in yearly revenue need that authorization.⁷⁶

Several MetaPrompt posts cover its records. MiniMax lists M3 at $1.20 per 1M output tokens for prompts up to 512K tokens, after a cut its pricing page labels “Permanent 50% off” Vendor-stated; in Cheapest LLM APIs per 1M Output Tokens (October 2026) that ties for sixth and is the cheapest top-tier row.⁸ Its earlier open-weight model MiniMax M2.5 resolved 75.8% of SWE-bench Verified tasks in the shared mini-SWE-agent v2.0.0 harness on February 17, 2026, tying Gemini 3 Flash, at an average cost of $0.07 per task Measured, as set out in SWE-bench Verified, Pro and What Replaced Them: By the Numbers.⁹ Its video model MiniMax H3 generates clips of up to 15 seconds at up to 2K resolution Vendor-stated, one of the rows in How AI Video Generation Works: Clip Length, Resolution and Audio by Model.¹⁰ On Hugging Face, MiniMax-M2.7 was its most-downloaded general language model on Oct 13, 2026, with 702,817 downloads in the previous 30 days Measured, a figure Chinese Open-Weight Models: Downloads, Derivatives and Token Share sets beside eight other Chinese labs.³