MiniMax M1: An Open Reasoning Model Built for Long Outputs

Reasoning models have a cost problem that is easy to miss on a pricing page. The expensive part is not reading your prompt; it is generating tens of thousands of thinking tokens before the answer, and with standard attention each of those tokens costs more than the last. MiniMax M1 was designed around that specific problem.

MiniMax announced M1 on 16 June 2025 as an open-weight reasoning model with a 1M-token context window. It sits between two other MiniMax models in time and in purpose: it is built on MiniMax-Text-01, and it was followed by the M2 line, which shifted the focus to agentic coding.

What M1 adds on top of Text-01

The architecture is inherited. M1 has 456B total parameters and 45.9B active per token, the same hybrid stack as MiniMax-01: most layers use lightning attention, a linear-attention variant, with a standard softmax attention layer at regular intervals. What M1 adds is a long reinforcement-learning phase that turns a general model into one that reasons step by step before answering.

MiniMax released two versions that differ in thinking budget, M1-40k and M1-80k, named for the maximum reasoning length they were trained to use. The 80k version generally scores a little higher, at the price of longer outputs.

The RL method is MiniMax's own, called CISPO. Rather than clipping individual token updates, as PPO-style methods do, it clips importance-sampling weights. The company reports that it converged about twice as fast as DAPO in its experiments, and that the full RL run took three weeks on 512 H800 GPUs at a rental cost of about $534,700. That is a vendor figure for one phase of training, not the total cost of the model.

The efficiency claim, and what it is measured against

MiniMax's central claim is that M1 needs about 25% of DeepSeek R1's FLOPs when generating 100K tokens, and about 30% at an 80K-token reasoning length. The reason is the attention mix: lightning attention's cost grows roughly linearly with length, so the gap widens the longer the model thinks.

Read it carefully. It is a compute comparison at long generation lengths, made by the vendor, against one specific model. At short outputs the advantage shrinks, and FLOPs are not the same as your bill, which depends on how a provider has priced and batched the model. It is a good reason to expect M1 to handle long reasoning gracefully, not proof that it will be cheaper for your workload.

Benchmarks: long context and tools, not coding

MiniMax's own table for the 80k and 40k versions shows a clear profile:

• AIME 2024: 86.0 and 83.3
• MATH-500: 96.8 and 96.0
• GPQA Diamond: 70.0 and 69.2
• LiveCodeBench: 65.0 and 62.3
• SWE-bench Verified: 56.0 and 55.6
• TAU-bench airline: 62.0 and 60.0
• OpenAI-MRCR at 128K: 73.4 and 76.1
• LongBench-v2: 61.5 and 61.0

MiniMax placed it behind only Gemini 2.5 Pro on long-context tasks at release, and ahead of other open-weight models. Software engineering was weaker: the SWE-bench Verified figures use an Agentless scaffold on a 486-task subset and sit well below the frontier of the time, and DeepSeek-R1-0528 scored slightly higher in the vendor's own comparison. If coding agents are your main use, M1 was never the strongest option, and the later M2 series was aimed squarely at coding and agent work.

Paying for thinking

MiniMax's launch pricing was tiered by input length: $0.40 per million input tokens up to 200K and $1.30 beyond that, with output at $2.20 throughout. At the time of writing OpenRouter lists M1 at $0.55 input and $2.20 output, with a 40,000-token output cap. The listing does not say whether the 40k or 80k weights sit behind it, so check with your provider which one you are getting.

Output is where the money goes. A reasoning model can spend many times more tokens thinking than answering, and every one of them bills at the output rate. The token cost of reasoning models works through the arithmetic, and reasoning models explained covers when longer thinking actually improves the answer.

Settings that matter

MiniMax recommends a temperature of 1.0 and top_p of 0.95. Lowering temperature, a common reflex for deterministic tasks, is not what the model was tuned for. For maths it suggests a system prompt asking it to reason step by step and put the final answer in \boxed{}, which makes answers easy to parse automatically.

The weights are published under Apache 2.0, a more permissive arrangement than MiniMax-01's custom model agreement. M1 supports function calling, and it runs on vLLM and Transformers. Self-hosting is still a large deployment, because all 456B parameters have to sit in memory even though only 45.9B are active.

M1, MiniMax-01 or something newer

Against its sibling the rule is simple. MiniMax-01 answers without visible reasoning and accepts images; M1 is text-only and reasons first. Use MiniMax-01 for long-document extraction and summaries, and M1 when the question needs several steps of work.

Against the newer line, M1 is mostly of interest for its licence, its long-output efficiency and its 1M window. The MiniMax M2 guide covers the model that replaced it for agentic work, with a much smaller active parameter count of about 10B. If your workload is long analytical reasoning over very large inputs, test M1 against it on your own tasks before assuming newer is better.