MiniMax-M1

2025-06-16 · Publisher : MiniMax · Domain : Artificial Intelligence

MiniMax releases MiniMax-M1, presented as the first large-scale open reasoning model with hybrid attention: 456 billion parameters, 1 million tokens of context, Apache 2.0 license.

### Context In June 2025, long reasoning, popularized by o1 and then by DeepSeek-R1 (open weights in January 2025), was still limited to 128,000-token contexts on the open side. On the closed side, Gemini 2.5 Pro already offered a million tokens and o3 dominated reasoning tests. MiniMax reuses its January 2025 architecture (MiniMax-Text-01) and adds reinforcement learning training on top of it. ### What the model brings - Long-context reasoning: the model supports 1 million tokens of input and up to 80,000 tokens of reasoning, i.e. 8 times the context of DeepSeek-R1 according to the model card. - Compute efficiency: thanks to Lightning Attention, the model card states that M1 consumes 25% of the FLOPs of DeepSeek-R1 for a 100,000-token generation. - CISPO: a new reinforcement learning algorithm that clips importance sampling weights rather than per-token updates, with more stable training on this type of MoE. - Documented training cost: the reinforcement learning phase used 512 H800 GPUs for three weeks, for $534,700 in rental according to the technical report. - Two variants: M1-40k and M1-80k (thinking budget), the 40k being an intermediate checkpoint. ### Architecture and specifications - Parameters: 456 billion in total, 45.9 billion active per token (same base as MiniMax-Text-01). - Context: 1 million native tokens; thinking budget of 40,000 or 80,000 tokens. - Modalities: text. Languages and knowledge cutoff: not disclosed in the sources consulted. - License: Apache 2.0 (text read in the GitHub repository), with no commercial-use restriction or user threshold. - Weights: https://huggingface.co/MiniMaxAI/MiniMax-M1-80k ### Cost and access | Item | Value | |---|---| | API, input 0 to 200,000 tokens | $0.4 per million tokens | | API, output (0 to 200,000 tokens) | $2.2 per million tokens | | API, input 200,000 to 1 million tokens | $1.3 per million tokens | | API, output (200,000 to 1 million) | $2.2 per million tokens | | Self-hosting | hardware not disclosed in the sources consulted | Pricing from the June 16, 2025 post. The post states that the first tier is cheaper than DeepSeek-R1 and that the second does not exist at DeepSeek. ### Benchmarks Scores published by the vendor in the MiniMax-M1-80k model card; thinking limits vary (80k for M1, 64k for Claude 4 Opus and Gemini 2.5 Pro, 100k for o3). | Test | MiniMax-M1-80k | OpenAI o3 | Gemini 2.5 Pro (06-05) | Claude 4 Opus | |---|---|---|---|---| | AIME 2024 | 86.0 | 91.6 | 92.0 | 76.0 | | GPQA Diamond | 70.0 | 83.3 | 86.4 | 79.6 | | LiveCodeBench | 65.0 | 75.8 | 77.1 | 56.6 | | SWE-bench Verified | 56.0 | 69.1 | 67.2 | 72.5 | | OpenAI-MRCR (128k) | 73.4 | 56.5 | 76.8 | 48.9 | | OpenAI-MRCR (1M) | 56.2 | not measured | 58.8 | not measured | | TAU-bench (airline) | 62.0 | 52.0 | 50.0 | 59.6 | Source: MiniMax-M1-80k model card (Hugging Face), vendor measurements. ### Versus closed models At the time of its release, the best closed reasoning models were o3 and Gemini 2.5 Pro. M1 remains behind them in mathematics (AIME 2024: 86.0 versus 91.6 and 92.0), in scientific reasoning (GPQA Diamond: 70.0 versus 83.3 and 86.4) and in software engineering (SWE-bench Verified: 56.0 versus 69.1 and 67.2). It comes close to Gemini 2.5 Pro on long context (OpenAI-MRCR at 1 million: 56.2 versus 58.8) and outperforms the three closed models cited on the TAU-bench airline tool-use test (62.0 versus 52.0, 50.0 and 59.6). On price, the vendor provides no numerical comparison with closed-model pricing, comparing itself instead with DeepSeek-R1. For the open versus closed choice, the Apache 2.0 license makes self-hosting and redistribution possible without permission, but a 456-billion-parameter model requires substantial GPU infrastructure. ### Key takeaways M1 brings to the open world reasoning over a million tokens under a permissive license, with a published reinforcement learning training cost. It does not catch up with o3 or Gemini 2.5 Pro on reasoning and coding tests, but it is usable for reasoning over very long documents in self-hosting.

Tags : LLM, Open Source, MiniMax, Raisonnement, MoE, Apache 2.0, Lightning Attention