MiniMax-M1

MiniMax-M1, released in June 2025, is the lab's first reasoning model and, by its own description, the world's first open-weight large-scale hybrid-attention reasoning model. It is not a new base model: it reuses MiniMax-Text-01 in full — the same 456 billion parameters, the same 45.9 billion active per token, the same lightning-attention stack — and adds large-scale reinforcement learning on top. The weight files of M1 and Text-01 are identical in size down to the byte. The reinforcement learning is where the work went. MiniMax introduced CISPO, an algorithm that clips importance-sampling weights rather than token updates, which the company reports as outperforming competing RL variants on the same budget. The hybrid attention pays off during long chains of thought: MiniMax reports that at a generation length of 100,000 tokens M1 consumes a quarter of the FLOPs of DeepSeek-R1, and it natively supports a one-million-token context — eight times R1's. M1 shipped in two variants distinguished only by thinking budget: 40K and 80K tokens of extended reasoning. More thinking is not uniformly better. In MiniMax's own table the 40K variant reads long context more accurately than the 80K one — 76.1 against 73.4 on OpenAI-MRCR at 128k, 58.6 against 56.2 at one million tokens — and it also wins the retail half of TAU-bench, 67.8 against 63.5. The figure most often quoted for M1, 73.4, is therefore the weaker of the two, and Gemini 2.5 Pro sits above both at 76.8. The comparison table is worth reading in both directions. MiniMax states that M1 outperforms strong open-weight models such as the original DeepSeek-R1, and against that January 2025 release the claim holds. The same table, however, carries a second DeepSeek column — the May 2025 R1-0528 — which beats M1-80K in 13 of the 16 rows both models report. M1's three wins are long-context reading (OpenAI-MRCR at 128k, LongBench-v2) and the airline half of TAU-bench. Two rows are also not like-for-like: the SWE-bench Verified score of 56.0 was produced by MiniMax's own two-stage Agentless scaffold on the 486 of 500 verified tasks that ran on its infrastructure, and the HLE scores carrying an asterisk — M1's included — come from the text-only subset, while the unmarked Claude, Gemini and o3 figures in the same row do not. M1's licence is the outlier in MiniMax's catalogue. Text-01 before it and every M2 release after it carry custom terms; M1 alone is Apache 2.0, which cannot be withdrawn. Independent providers still host it at roughly $0.55 per million input tokens and $2.20 per million output — today more expensive than MiniMax's own newer and larger-context M3.

▸Videos

No videos yet.

▸Reviews