Overview
MiniMax is a Chinese AI research lab and product company whose model family spans language, video, speech, image, and music. Its current text flagship, MiniMax M3, is built on a novel sparse-attention architecture (MSA) that reaches a 1-million-token context and is positioned for coding and agentic workflows; the company also ships MiniMaxCode, an agent harness that assembles skill teams to solve longer tasks. On the generative side, MiniMax H3 is an open-weight omni-modal model and Hailuo remains a popular text-to-video line. In our evaluation, MiniMax’s differentiator is the long context plus multimodal breadth from one vendor: you can run document analysis, agents, and video generation inside one account. It is a strong choice for teams already operating in the China ecosystem or wanting an alternative to the usual US frontier models. The caveats are regional payment and access friction, plus API pricing that has moved around, so cost modeling should be done against the live rate card.
Key Features
- MiniMax M3 language model with 1M-token context for coding and long-document agents
- MiniMaxCode agent harness that learns habits and assembles skill teams
- Open-weight H3 omni-modal and Hailuo text-to-video generation
- Speech-2.8 HD/Turbo for natural multilingual text-to-speech
- Both Anthropic-compatible and OpenAI-compatible API endpoints
- Self-hostable open-weight models (M3, H3) for on-prem control
Pricing
| Plan | Price | For | Notes |
|---|---|---|---|
| Free credits | $0 | New API accounts | Trial allowance; no music paid API for new users |
| Individual plan | $49/mo | Solo developers | Subscription after 2026 repricing (was $29) |
| M3 API <=512K ctx | $4.20 in / $16.80 out / 1M tok | Pay-as-you-go text | Standard rate; ~$0.59 / $2.37 per 1M (approx 7.1 RMB/USD) |
| M3 API 512K-1M ctx | $8.40 in / $33.60 out / 1M tok | Long-context calls | Premium band for 1M-context workloads |
| Open-weight self-host | Free | Orgs wanting control | M3/H3 weights on Hugging Face and ModelScope |
Comparison
Next to DeepSeek, MiniMax prices higher but leans into long context and multimodal breadth rather than pure cost leadership. Against Qwen, both are Chinese open-weight players, but MiniMax pushes a 1M-token coding model and a separate agent harness. Versus Kimi, which is best known for long-text chat, MiniMax covers more modalities (video, speech, music) under one roof. Choose MiniMax when you want long-context agents or multi-modal generation from a single vendor.