# MiniMax-M1 (M1-40k / M1-80k)

> The first open-weight, large-scale hybrid-attention reasoning model — 456B-param MoE with a native 1M-token context.

- Maker: MiniMax
- Line: MiniMax M1
- Canonical page: https://ai-tldr.dev/models/minimax-m1/
- Updated: 2026-06-17

## In short

MiniMax-M1 (M1-40k / M1-80k) is the first open-weight, large-scale hybrid-attention reasoning model — 456B-param MoE with a native 1M-token context. MiniMax released it on 16 June 2025. It has a 1M-token context window. You can use it through MiniMax and OpenRouter.

## Facts

| Fact | Value | Source |
| --- | --- | --- |
| Context window | 1M (max output 80k tokens (M1-80k); 40k tokens (M1-40k)) | [1] |
| Input price | $0.40 / 1M tokens (0-200k context); $1.30 / 1M tokens (200k-1M context) | [2] |
| Output price | $2.20 / 1M tokens | [2] |
| Licence | Apache 2.0 · open weights | [1] |
| Released | 16 June 2025 | [3] |
| Parameters | 456B total / 45.9B active (MoE, 32 experts) | [1] |
| Knowledge cutoff | June 2024 | [1] |

## Sources

1. [Hugging Face model card (M1-80k)](https://huggingface.co/MiniMaxAI/MiniMax-M1-80k)
2. [MiniMax-M1 announcement](https://www.minimax.io/news/minimaxm1)
3. [GitHub repository](https://github.com/MiniMax-AI/MiniMax-M1)

## Overview

MiniMax-M1 is, per MiniMax, the world's first open-weight, large-scale hybrid-attention reasoning model. It is built on the MiniMax-Text-01 base and uses a Mixture-of-Experts design with 456 billion total parameters, of which 45.9 billion are activated per token across 32 experts. Its defining trait is a hybrid attention stack: a softmax-attention transformer block follows every seven 'lightning' (linear) attention blocks, which gives near-linear cost as the sequence grows. MiniMax reports that at a 100K-token generation length M1 uses roughly 25% of the FLOPs that DeepSeek-R1 would.

M1 ships in two variants that differ only in their reasoning (thinking) budget: MiniMax-M1-40k and MiniMax-M1-80k. Both natively support a 1-million-token context window — eight times that of DeepSeek-R1 and on par with closed models like Gemini 2.5 Pro. The 80k variant generally scores a little higher on reasoning and coding, while the 40k variant is cheaper to run and actually leads on some long-context and agent benchmarks. Both were trained with reinforcement learning using MiniMax's CISPO algorithm, which clips importance-sampling weights rather than token updates; the full RL run took 512 H800 GPUs about three weeks at a reported rental cost of $534,700.

The weights are released under the permissive Apache 2.0 license on Hugging Face and GitHub, and MiniMax recommends vLLM (0.9.2+) or Transformers for deployment. M1 is text-only. It is positioned for long-context understanding, software-engineering tasks, and agentic tool use, where MiniMax reports it tops other open-weight models and is competitive with leading proprietary systems.

## Benchmarks

- AIME 2024: 86% ([source](https://arxiv.org/html/2506.13585v1))
- AIME 2024 (M1-40k): 83.3% ([source](https://arxiv.org/html/2506.13585v1))
- AIME 2025: 76.9% ([source](https://arxiv.org/html/2506.13585v1))
- LiveCodeBench: 65% ([source](https://arxiv.org/html/2506.13585v1))
- SWE-bench Verified: 56% ([source](https://arxiv.org/html/2506.13585v1))
- SWE-bench Verified (M1-40k): 55.6% ([source](https://arxiv.org/html/2506.13585v1))
- MMLU-Pro: 81.1% ([source](https://arxiv.org/html/2506.13585v1))
- GPQA Diamond: 70% ([source](https://arxiv.org/html/2506.13585v1))
- TAU-bench (airline, M1-80k): 62% ([source](https://arxiv.org/html/2506.13585v1))
- TAU-bench (retail, M1-40k): 67.8% ([source](https://arxiv.org/html/2506.13585v1))
- OpenAI-MRCR (1M, M1-40k): 58.6% ([source](https://arxiv.org/html/2506.13585v1))
- LongBench-v2: 61.5% ([source](https://arxiv.org/html/2506.13585v1))

## Pricing

- Input: $0.40 / 1M tokens (0-200k context); $1.30 / 1M tokens (200k-1M context) per 1M tokens
- Output: $2.20 / 1M tokens per 1M tokens

Tiered input pricing by context length from MiniMax's official announcement; OpenRouter lists $0.40 in / $2.20 out. Free unlimited use is offered on the MiniMax app and web.

Source: https://www.minimax.io/news/minimaxm1

## Strengths

- Native 1M-token context — among the largest of any open-weight model, matching closed frontier models
- Efficient long-form reasoning: lightning attention cuts FLOPs to ~25% of DeepSeek-R1 at 100K-token generation
- Strong agentic tool use — leads open-weight models on TAU-bench and beats Gemini 2.5 Pro on parts of it
- Apache 2.0 license with fully open weights — free commercial use and self-hosting
- Two thinking budgets (40k/80k) let you trade reasoning depth against cost
- Competitive coding and math (AIME 2024 86.0%, SWE-bench Verified 56.0% on the 80k variant)

## Best for

- Long-document and whole-codebase analysis that needs hundreds of thousands of tokens of context
- Agentic tool-use and function-calling workflows
- Software engineering: bug fixing and repo-level tasks (SWE-bench-style)
- Math and competition-style reasoning
- Self-hosted reasoning deployments where an open Apache-2.0 license is required

## How to access

- [MiniMax](https://www.minimax.io/platform_overview) — model id `MiniMax-M1`
- [OpenRouter](https://openrouter.ai/minimax/minimax-m1) — model id `minimax/minimax-m1`

## MiniMax M1 versions

- MiniMax-M1 (M1-40k / M1-80k) (June 2025) — this page

## FAQ

### What is the difference between MiniMax-M1-40k and MiniMax-M1-80k?

They are the same 456B-parameter model trained with two different reasoning (thinking) budgets: 40,000 tokens versus 80,000 tokens. The 80k variant generally scores slightly higher on reasoning and coding benchmarks, while the cheaper 40k variant leads on some long-context and agentic tool-use tasks. Both share the same 1M-token context window.

### Is MiniMax-M1 open source and free to use?

The weights are released under the Apache 2.0 license on Hugging Face and GitHub, so you can download, self-host, and use them commercially for free. MiniMax also offers a hosted API (tiered pricing) and free unlimited use through its own app and website.

### How large is the context window?

MiniMax-M1 natively supports a 1-million-token context window — about eight times DeepSeek-R1's and comparable to closed frontier models. Its lightning (linear) attention makes processing very long inputs far cheaper than standard softmax attention.

### Does MiniMax-M1 support images or audio?

No. MiniMax-M1 is a text-only reasoning model. It is designed for long-context text understanding, coding, math, and agentic tool use rather than multimodal input.

## Links

- [MiniMax-M1 announcement](https://www.minimax.io/news/minimaxm1)
- [GitHub repository](https://github.com/MiniMax-AI/MiniMax-M1)
- [Hugging Face model card (M1-80k)](https://huggingface.co/MiniMaxAI/MiniMax-M1-80k)
- [Technical report (arXiv 2506.13585)](https://arxiv.org/abs/2506.13585)
- [MiniMax API platform](https://www.minimax.io/platform_overview)
- [OpenRouter (MiniMax M1)](https://openrouter.ai/minimax/minimax-m1)
- [OpenRouter benchmarks](https://openrouter.ai/minimax/minimax-m1/benchmarks)
- [MiniMax-M1 (M1-40k / M1-80k) benchmarks](https://arxiv.org/html/2506.13585v1)

Source: AI/TLDR LLM registry, https://ai-tldr.dev/models/minimax-m1/