MiniMax
Chinese multimodal models: text, voice and video.
Olya's profile
- What sets it apart
- MiniMax is a Chinese lab that keeps an entire multimodal stack under one roof: the M-series text models (up to M3, with MSA sparse attention and context up to 1M tokens), Hailuo video generation, Speech voice synthesis and music. Its real distinguishing trait is that some models, starting with M2, ship with open weights under an MIT license, so you can download them and use them commercially.
- Strengths
- Its per-token price is among the lowest at the frontier: the M-series lists at roughly $0.30 per million input tokens and $1.20 output on the official sheet. M3 is tuned for coding and agentic workflows, Hailuo is competitive on generated video, and the voices are realistic and multilingual. If you want control, you can self-host the open models instead of relying on the API.
- When to use it
- It pays off when you want the best cost/performance ratio on agents, coding assistance, or high-volume video and voice generation, where Western pricing becomes prohibitive. It's also a good choice if you want to start from a competitive open model and run it on your own servers.
- When to avoid it
- If you handle sensitive data or face compliance constraints, weigh the jurisdiction carefully: it's a Chinese lab, and its terms of service and Western availability shift often. For top-tier general reasoning or a more mature ecosystem of docs and integrations, the frontier models from OpenAI, Anthropic or Google are still ahead.