Superseded (verified 2026-09-30)

M1 (June 2025) is two generations old. MiniMax’s current open-weight line is M2.x and minimax-m3 (2026).

MiniMax M1

MiniMax’s first open-weight reasoning model, released June 2025 during its “MiniMaxWeek”.

Specifications (Hugging Face card, M1-80k)

• Parameters: 456B total / 45.9B active per token (hybrid MoE)
• Context: native 1M tokens, with “lightning attention”
• Licence: Apache 2.0 (earlier note only said “open-weight”)
• Reported results: SWE-bench Verified 56.0%; OpenAI-MRCR at 1M tokens 56.2%. The card says it outperforms the original DeepSeek-R1 and Qwen3-235B on software engineering, tool use and long-context tasks (vendor claim).
• Serving: vLLM 0.9.2+ recommended

Claims carried over, unverified this session

• “75% less compute than DeepSeek R1” and the CISPO training method (half the training steps of comparable approaches) come from MiniMax’s launch material; not re-checked.

See also

• MiniMax-Text-01, MiniMax M2.x, minimax-m3
• MiniMax-VL-01, MiniMax Speech-02

Sources

• https://huggingface.co/MiniMaxAI/MiniMax-M1-80k (accessed 2026-09-30)