We introduce the MiniMax-M2 series, a family of Mixture-of-Experts language models built around the principle that mini activations can unleash maximum real-world intelligence. The flagship M2 contains 229.9B total parameters with only 9.8B activated per token. Designed end-to-end for agentic deployment, the M2 series rests on three components: (i) agent-driven data pipelines producing large-scale, verifiable trajectories across agentic coding and agentic cowork, each grounded in an executable workspace and an artifact-aligned reward; (ii) Forge, a scalable agent-native RL system that adapts to long-horizon agent trajectories, paired with windowed-FIFO scheduling, prefix-tree merging, inference optimization, and a clean training-inference-agent decoupling that supports both white-box and black-box agents; (iii) the latest M2.7 checkpoint takes an early step toward self-evolution -- autonomously debugging training runs and modifying its own scaffold. Across M2 through M2.7, this combination translates a mini-activation footprint into frontier-tier performance on agentic coding, deep search, office-task, and reasoning benchmarks.
核心贡献 · Key contributions
提出 MiniMax-M2 系列，MoE 模型总参数量 229.9B，每 token 激活 9.8B 参数，达到前沿性能。 Introduces MiniMax-M2 series, a MoE model with 229.9B total and 9.8B activated parameters, achieving frontier performance.
设计智能体驱动的数据流水线，覆盖编码、协作、推理和通用任务，提供可验证奖励。 Designs agent-driven data pipelines for coding, cowork, reasoning, and general tasks with verifiable rewards.
构建 Forge，可扩展的智能体原生强化学习系统，采用窗口 FIFO 调度和前缀树合并。 Builds Forge, a scalable agent-native RL system with windowed-FIFO scheduling and prefix-tree merging.
在 M2.7 中展示自我进化能力，自主调试训练运行并修改自身框架。 Demonstrates self-evolution in M2.7, autonomously debugging training runs and modifying its own scaffold.
以极小激活量在智能体编码、深度搜索、办公任务和推理基准上取得强劲结果。 Achieves strong results on agentic coding, deep search, office-task, and reasoning benchmarks with mini activations.
提出混合领域 RL 训练，避免跨领域的灾难性遗忘和负迁移。 Proposes mixed-domain RL training to avoid catastrophic forgetting and negative transfer across domains.
局限 · Limitations
混合注意力变体在长上下文任务上性能劣于全注意力。 Hybrid attention variants showed degraded performance on long-context tasks compared to full attention.
自我进化仅限于早期操作形式，可能无法泛化到所有任务。 Self-evolution is limited to early operational form and may not generalize to all tasks.
智能体驱动数据流水线依赖高质量种子数据，大规模下可能引入噪声。 Agent-driven data pipelines rely on high-quality seed data and may introduce noise at scale.
RL 训练稳定性依赖于领域混合比例和难度分布的精细调优。 RL training stability depends on careful tuning of domain mixing ratios and difficulty distribution.
模型在非英语或低资源语言上的性能未得到广泛评估。 The model's performance on non-English or low-resource languages is not extensively evaluated.
论文章节 · Sections（共 21）
摘要Abstract
1 引言1 Introduction
2.1 整体架构2.1 Overall Architecture
2.2 模型设计选择2.2 Model Design Choice
2.3 多令牌预测2.3 Multi-Token Prediction
3 预训练数据3 Pre-Training Data
4.1 智能体编码4.1 Agentic Coding
4.2 智能体协作4.2 Agentic Cowork
4.3 推理密集型任务4.3 Reasoning-Intensive Tasks
4.4 通用对话与写作4.4 General-Purpose Conversation and Writing
4.5 角色扮演与人格一致性4.5 Role-Play and Persona Coherence
5 监督微调5 Supervised Fine-Tuning
6.1 强化学习算法6.1 RL Algorithm
6.2 强化学习基础设施6.2 RL Infrastructure
7.1 交错思考7.1 Interleaved Thinking
7.2 自我进化7.2 Self-Evolution
8 评估8 Evaluation
8.1 评估设置8.1 Evaluation Settings
8.2 主要结果8.2 Main Results
8.3 案例研究：MLE Bench Lite 上的自我进化8.3 Case Study: Self-Evolution on MLE Bench Lite