MiniMax-M1：利用闪电注意力高效扩展测试时计算

MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention

闫俊杰 Junjie Yan · MiniMax · 2025-06-16 · arXiv:2506.13585 ↗ · 被引 185

摘要 · Abstract

我们推出了 MiniMax-M1，这是全球首个开源的大规模混合注意力推理模型。MiniMax-M1 采用混合专家（MoE）架构与闪电注意力机制，基于之前的 MiniMax-Text-01 模型开发，总参数量 4560 亿，每个 token 激活 459 亿参数。M1 原生支持 100 万 token 的上下文长度，是 DeepSeek R1 的 8 倍。此外，MiniMax-M1 中的闪电注意力机制能够高效扩展测试时计算，使其特别适合需要处理长输入和深度思考的复杂任务。MiniMax-M1 通过大规模强化学习（RL）在包括沙盒环境和真实软件工程环境在内的多样化问题上进行训练。除了 M1 在 RL 训练中固有的效率优势外，我们提出了 CISPO，一种新的 RL 算法，通过裁剪重要性采样权重而非 token 更新，进一步提升了 RL 效率，优于其他竞争性 RL 变体。结合混合注意力和 CISPO，MiniMax-M1 的完整 RL 训练仅需在 512 块 H800 GPU 上运行三周，租赁成本仅为 53.47 万美元。我们发布了两个版本的 MiniMax-M1 模型，分别具有 40K 和 80K 的思考预算，其中 40K 模型是 80K 训练的中间阶段。在标准基准测试中，我们的模型与 DeepSeek-R1 和 Qwen3-235B 等强开源模型相当或更优，在复杂软件工程、工具使用和长上下文任务中表现尤为突出。我们已在 https://github.com/MiniMax-AI/MiniMax-M1 公开发布 MiniMax-M1。

We introduce MiniMax-M1, the world's first open-weight, large-scale hybrid-attention reasoning model. MiniMax-M1 is powered by a hybrid Mixture-of-Experts (MoE) architecture combined with a lightning attention mechanism. The model is developed based on our previous MiniMax-Text-01 model, which contains a total of 456 billion parameters with 45.9 billion parameters activated per token. The M1 model natively supports a context length of 1 million tokens, 8x the context size of DeepSeek R1. Furthermore, the lightning attention mechanism in MiniMax-M1 enables efficient scaling of test-time compute. These properties make M1 particularly suitable for complex tasks that require processing long inputs and thinking extensively. MiniMax-M1 is trained using large-scale reinforcement learning (RL) on diverse problems including sandbox-based, real-world software engineering environments. In addition to M1's inherent efficiency advantage for RL training, we propose CISPO, a novel RL algorithm to further enhance RL efficiency. CISPO clips importance sampling weights rather than token updates, outperforming other competitive RL variants. Combining hybrid-attention and CISPO enables MiniMax-M1's full RL training on 512 H800 GPUs to complete in only three weeks, with a rental cost of just $534,700. We release two versions of MiniMax-M1 models with 40K and 80K thinking budgets respectively, where the 40K model represents an intermediate phase of the 80K training. Experiments on standard benchmarks show that our models are comparable or superior to strong open-weight models such as the original DeepSeek-R1 and Qwen3-235B, with particular strengths in complex software engineering, tool utilization, and long-context tasks. We publicly release MiniMax-M1 at https://github.com/MiniMax-AI/MiniMax-M1.

核心贡献 · Key contributions

首个开源的大规模混合注意力推理模型，采用闪电注意力机制实现高效的测试时算力扩展。
First open-weight large-scale hybrid-attention reasoning model with lightning attention for efficient test-time compute scaling.
提出 CISPO 算法，裁剪重要性采样权重而非词元更新，相比 DAPO 实现 2 倍加速。
Proposes CISPO, a novel RL algorithm clipping importance sampling weights instead of token updates, achieving 2x speedup over DAPO.
原生支持 100 万词元的上下文长度，是 DeepSeek R1 的 8 倍，在 10 万词元生成长度下仅消耗 25%的 FLOPs。
Achieves 1M token native context length, 8x DeepSeek R1, with 25% FLOPs at 100K generation length.
在 512 块 H800 GPU 上完成完整强化学习训练仅需 3 周，租赁成本 53.47 万美元。
Full RL training on 512 H800 GPUs completes in 3 weeks with $534,700 rental cost.
在 SWE-bench Verified、TAU-bench 和长上下文基准上超越开源模型，在工具使用上超过 Gemini 2.5 Pro。
Outperforms open-weight models on SWE-bench Verified, TAU-bench, and long-context benchmarks, surpassing Gemini 2.5 Pro on tool use.
引入基于沙盒的强化学习环境用于真实软件工程，实现基于执行的奖励。
Introduces sandbox-based RL environments for real-world software engineering, enabling execution-based rewards.

局限 · Limitations

在数学和编程竞赛上落后于 DeepSeek-R1-0528。
Lags behind DeepSeek-R1-0528 on mathematical and coding competitions.
强化学习训练需要谨慎处理精度不匹配和优化器超参数问题。
RL training requires careful handling of precision mismatch and optimizer hyperparameters.
通用强化学习数据整理和奖励模型校准复杂且劳动密集。
General RL data curation and reward model calibration are complex and labor-intensive.
长上下文强化学习扩展面临模式崩溃和负样本不平衡问题。
Long-context RL scaling faces pattern collapse and negative sample imbalance issues.
在 SimpleQA 上的事实性表现不如 DeepSeek-R1。
Factuality on SimpleQA underperforms DeepSeek-R1.

论文章节 · Sections（共 16）

摘要 Abstract
1 引言 1 Introduction
2 可扩展强化学习准备：持续预训练与 SFT 2 Preparation for Scalable RL: Continual Pretraining and SFT
2.1 持续预训练：强化学习扩展的基础 2.1 Continual Pre-Training: Foundation for RL Scaling
2.2 监督微调：面向高效强化学习的聚焦对齐 2.2 Supervised Fine-Tuning: Focused Alignment for Efficient RL
3 高效强化学习扩展：算法与闪电注意力 3 Efficient RL Scaling: Algorithms and Lightning Attention
3.1 基于 CISPO 的高效强化学习扩展 3.1 Efficient RL Scaling with CISPO
3.2 基于闪电注意力的高效强化学习扩展：挑战与策略 3.2 Efficient RL Scaling with Lightning Attention – Challenges and Recipes
4 多样化数据驱动的强化学习扩展 4 Scaling Reinforcement Learning with Diverse Data
4.1 基于规则验证的推理密集型任务 4.1 Reasoning-Intensive Tasks with Rule-based Verification
4.2 基于模型反馈的通用领域任务 4.2 General Domain Tasks with Model-based Feedbacks
4.3 多样化数据融合的课程学习 4.3 Curriculum of Incorporating Diverse Data
5 将强化学习扩展至更长的思考过程 5 Extending RL Scaling to Longer Thinking
6.1 核心基准测试 6.1 Core Benchmarks
6.2 强化学习扩展的效果 6.2 Effect of RL Scaling
7 结论与未来工作 7 Conclusion and Future work

阅读逐段中英对照全文 →