PaperlayerView exact revision on arXiv

Source-linked AI summary

The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence

Aili Chen, Aonian Li, Baichuan Zhou, Bangwei Gong, Binyang Jiang, Boji Dan, Changhao Zhang, Changqing Yu, Chao Wang, Cheng Ma, Cheng Zhong, Cheng Zhu, Chengjun Xiao, Chengyi Yang, Chengyu Du, Chenyang Zhang, Chi Zhang, Chuangyi Huang, Chunhao Zhang, Chunhui Du, Chunyu Zhao, Congchao Guo, Da Chen, Deming Ding, Dianjun Sun, Dong Li, Dongyu Zhang, Enhui Yang, Fei Yu, Guang Zheng, Guodong Zheng, Guohong Li, Haichao Zhu, Haigang Zhou, Haimo Zhang, Han Ding, Hao Zhang, Haohai Sun, Haolin Lyu, Haonan Lu, Haoyu Wang, Huajie Shi, Huiyang Li, Jiacheng Chen, Jian Zhang, Jiaqi Zhuang, Jiaren Cai, Jiaxin Pan, Jiayao Li, Jiayuan Song, Jichuan Zhang, Jie Wang, Jihao Gu, Jin Zhu, Jingwei Dong, Jingyang Li, Jingyu Zhang, Jingze Zhuang, Jinhao Tian, Jinli Liu, Jinyi Hu, Jun Tao, Jun Zhang, Junbin Ruan, Junhao Xu, Junjie Yan, Junteng Liu, Junxian He, Kang Xu, Ke Ji, Ke Yang, Kecheng Xiao, Keyu Duan, Keyu Li, Le Han, Letian Ruan, Li Yuan, Lianfei Yu, Liheng Feng, Lijie Mo, Lin Li, Linge Du, Lingye Bao, Lingyu Yang, Lingyuan Zhou, Loki, Lu Chen, Lunbin Zeng, Ming Li, Ming Zhong, Mingliang Tao, Mingyuan Chi, Mujie Lin, Nan Hu, Ningxin Chen, Peiyin Zhu, Peng Gao, Pengcheng Gao, Pengfei Li, Penglin Li, Pengyu Zhao, Qibin Ren, Qibing Ren, Qidi Xu, Qihan Ren, Qile Li, Qin Wang, Quanliang Chen, Qunhong Zeng, Rong Tian, Rongxin Guo, Rui Dong, Ruitao Leng, Ruize Zhang, Shanqi Liu, Shaoxiang Chen, Shaoyu Chen, Sheng Jia, Shun Yao, Shuoran Zhao, Shuqi Yu, Sichen Li, Sicheng Pan, Songquan Zhu, Tengfei Li, Tian Xie, Tiancheng Qin, Tianle Li, Tianrun Liang, Wei Liu, Weiqi Xu, Weitao Li, Weixiang Chen, Weiyu Cheng, Weiyu Zhang, Wenhu Chen, Wenqian Zhao, Xiancai Chen, Xiangjun Song, Xiangyuan Wang, Xianzhen Luo, Xiao Luo, Xiao Su, Xiaobo Li, Xiaodong Han, Xiaojie Wu, Xihao Song, Xingyi Han, Xinyu Guan, Xuan Lu, Xun Zou, Xunhao Lai, Xutong Li, Xuyang Shen, Yan Gong, Yan Ma, Yang Jiao, Yang Wang, Yang Xu, Yangsen Wang, Ye Tang, Yicheng Chen, Yihang Wang, Yinran Qiu, Yiqi Shi, Yiting Guo, Yiwen Huang, Yixuan Wang, Yongyi Hu, Yu Gao, Yu Zhang, Yuan Li, Yuanxiang Ying, Yuanzhen Zhang, Yubo Wang, Yuchen Song, Yufeng Yang, Yuhang Meng, Yuhang Miao, Yuhao Li, Yujie Liu, Yulin Hu, Yunan Huang, Yunji Li, Yunyi Huang, Yusen Zhang, Yusu Hong, Yutao Xie, Yutong Zhang, Yuwen Liao, Yuxuan Shi, Yuze Wenren, Zebin Li, Zehan Li, Zejian Luo, Zeyu Jin, Zeyuan Sun, Zhanpeng Zhou, Zhaochen Su, Zhendong Li, Zhengmao Zhu, Zhengyuan Peng, Zhenhua Fan, Zhi Zhang, Zhichao Xu, Zhiheng Lv, Zhikang Xu, Zhitao He, Zhiwei He, Zhongyuan Li, Zibo Gao, Zijia Wu, Zijian Song, Zijian Zhou, Zijun Sun, Zishan Huang, Ziying Chen, Ziyue Ge

arXiv:2605.26494v2cs.AIcs.CLcs.LG

TL;DR

Long-horizon agentic workflows create major efficiency and cost bottlenecks while demanding complex real-world task performance. The MiniMax-M2 series combines sparse MoE models, agent-driven data, Forge reinforcement learning, and early self-evolution, achieving parity with frontier systems an order of magnitude larger in per-step compute across agentic coding, cowork, and reasoning benchmarks.

  • Problem

    Long-horizon agentic workflows require costly, efficient training and inference while handling intrinsically complex real-world tasks.

  • Method

    The M2 series combines sparse MoE models with executable-workspace data pipelines, Forge agent-native reinforcement learning, and early self-evolution.

  • Results

    M2.7 achieves parity with frontier systems an order of magnitude larger in per-step compute across agentic coding, agentic cowork, and reasoning and knowledge benchmarks.

  • Takeaways & Limitations

    A roughly 10B activated-parameter footprint can match frontier systems with substantially larger per-step compute across multiple agentic capability areas.

  • Takeaways & Limitations

    Proxy metrics for complex multi-hop reasoning correlate fragilely with downstream performance, while statistically significant evaluation becomes substantially more compute-intensive as tasks grow complex.

Abstract

from arXiv · show

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.

1. Introduction

The MiniMax-M2 series applies mini activations to agentic intelligence through specialized data pipelines, the Forge reinforcement-learning system, and continued capability evolution. Its latest M2.7 checkpoint reports strong results across agentic coding, cowork, reasoning, and knowledge benchmarks with approximately 10B activated parameters.

  • Design principle: MiniMax-M2 is a Mixture-of-Experts family built around mini activations, with flagship M2 using 229.9B total parameters and 9.8B activated per token.The flagship is a 62-layer decoder-only Transformer organized as 256 fine-grained experts.
  • Technical innovations: The series uses high-fidelity agent data pipelines with executable environments, verifiable rewards, or credible feedback for coding, cowork, reasoning, and knowledge tasks.The introduction emphasizes reward quality and credibility for accepted trajectories, including executable verification and judge-model evidence checking.
  • Technical innovations: Forge is an agent-native reinforcement-learning system that unifies white-box and black-box agents while decoupling training, inference, and the agent itself.The system is designed for large-scale, general-purpose agentic reinforcement learning.
  • Results: 56.2 on SWE-bench Pro, 76.5 on SWE-bench Multilingual, 52.7 on Multi-SWE-bench, and 57.0 on Terminal-Bench 2.0 are reported for M2.7 agentic coding.These are the headline coding results previewed in Figure 1.
  • Results: 62.7 on MM Claw, 77.8 on BrowseComp, 50.0 on GDPval-AA, 46.3 on Toolathlon, 94.2 on AIME 2026, and 89.8 on GPQA-Diamond are reported for M2.7.The cited results span agentic cowork, reasoning, and knowledge capabilities.

2. Pre-Training Architecture

M2 is a 229.9B-parameter sparse MoE Transformer that activates 9.8B parameters per token, combining full multi-head attention, modified expert routing, and Multi-Token Prediction. Full attention was selected after hybrid alternatives degraded retrieval, multi-hop reasoning, in-context learning, and long-context performance.

  • Core architecture: M2 contains 229.9B total parameters with 9.8B activated per token in a 62-layer decoder-only Transformer.It uses hidden dimension 3,072, vocabulary size 200,064, 29.2T pre-training tokens, and a 192K maximum context length.
  • Attention and MoE design: Each block combines full multi-head attention with an MoE feed-forward layer, using 48 query heads, 8 key-value heads, and RoPE throughout.The model uses grouped-query attention while retaining full attention across all layers.
  • Attention and MoE design: M2’s MoE feed-forward layers use fine-grained experts, sigmoid gating, and learnable expert biases to increase routing diversity, smooth dynamics, and reduce load-balancing requirements.Fine-grained experts increase the number of smaller experts; sigmoid gating removes softmax’s zero-sum constraint; biases regulate expert utilization.
  • Attention selection: Hybrid SWA variants degraded retrieval, multi-hop reasoning, and in-context learning, with significantly worse performance than full attention on benchmarks exceeding 32K context after SFT.Within 32K context, differences were mixed and small in absolute terms, but long-context agent and complex-evaluation gaps became more pronounced.
  • Multi-Token Prediction: MTP jointly predicts the next K tokens, improving benchmark performance and enabling speculative decoding while preserving identical output quality to standard autoregressive decoding.Pre-training uses one MTP module (K = 1), with loss weight annealed from 0.3 to 0.1; continued pre-training expands to three modules (K = 3) via weight copying.

3. Pre-Training Data

The pre-training corpus combines quality-filtered, diverse sources with balanced domain sampling that upweights code, mathematics, and STEM. Training uses 19.9T tokens during constant pre-training, followed by a 9.3T-token decay phase extending context from 8K to 192K tokens.

  • Training Data: The corpus spans web documents, academic literature, books, programming code, and structured question-answering content, filtered by model-based rewards and auxiliary classifiers.Balanced sampling upweights high-quality content while retaining category diversity.
  • Data Distribution: Code, mathematics, and STEM are significantly upsampled, while general web content, books, and domain-specific data preserve broad world-knowledge and linguistic coverage.During the constant phase of pre-training, the dataset contains 19.9T tokens.
  • Long-Context Extension: The context window grows from 8K tokens through 32K to 192K tokens using a multi-stage procedure after initial pre-training.The decay phase uses a total data budget of 9.3T tokens.
  • Long-Context Extension: Long-context samples primarily use high-quality code concatenation, naturally long-form PDF documents, and thematically related document packing.High-quality data is mixed into decay training to consolidate capabilities while extending effective context length.

4. Post-Training Data Collection · 4.1. Agentic Coding

The agentic-coding data pipeline spans software engineering, application development, and terminal interaction, combining executable environments with task-specific verification. Its SWE, AppDev, and Terminal-Gym components scale diverse, verifiable training data through repository mining, expert synthesis, interactive application evaluation, and automated environment generation.

  • 4.1. Agentic Coding: Agentic-coding data covers software engineering, application development, and terminal interaction tasks across repository evolution, full-stack development, and interactive terminal environments.The three domains provide complementary coverage for coding-agent post-training.
  • 4.1.1. Real-Data Driven Collection: Software Engineering Tasks: SWE data combines problem statements, test-based verifiable rewards, and runnable Docker environments across more than ten programming languages and diverse coding categories.The pipeline filters GitHub pull requests, constructs executable environments, and routes task types to distinct reward formulations.
  • 4.1.1. Real-Data Driven Collection: Software Engineering Tasks: An agent-driven execution loop iteratively generates and refines build scripts using execution feedback to improve environment synthesis, especially for non-Python projects.The loop addresses heterogeneous dependencies, version conflicts, and repository-level variation.
  • 4.1.1. Real-Data Driven Collection: Software Engineering Tasks: SWE task diversification includes bug injection, commit merging, SWE-Test conversion, and code review, while model validation produces self-contained executable specifications.These transformations expand repair complexity, test-writing coverage, and task diversity while preserving verification.
  • 4.1.2. Expert-Driven Data Collection: Application Development Tasks (AppDev): AppDev uses expert-designed query synthesis, system prompts, and rejection sampling to generate complete applications evaluated beyond static code analysis.The pipeline combines domain expertise with automated quality control and trajectory sampling.
  • 4.1.2. Expert-Driven Data Collection: Application Development Tasks (AppDev): Agent-as-a-Verifier evaluates running applications through execution, interaction, and rubric-based assessment, using overall pass rate for rejection sampling and execution checks as hard gates.The verifier actively interacts with applications across multiple turns rather than judging static code or screenshots.
  • 4.1.3. Terminal-Gym: Automated Task Synthesis and Environment Generation: Terminal-Gym converts filtered Stack Overflow programming and system-operation scenarios into structured, tiered, Dockerized terminal tasks with generated tests and iterative diagnostic repair.The framework supports scalable terminal-operation training and is being extended toward Anything2Docker and broader cybersecurity research.

4.2. Agentic Cowork … 4.5. Role-Play and Persona Coherence

The paper builds Agentic Cowork and complementary reasoning, conversation, and role-play data tracks to support grounded professional work, diverse problem solving, general-purpose interaction, and coherent long-horizon dialogue. These pipelines emphasize executable or verifiable tasks, broad coverage, quality control, and explicit coherence objectives.

  • 4.2. Agentic Cowork: Agentic Cowork spans deep search, office deliverables, financial analysis and spreadsheets, and slide generation and editing for heterogeneous professional environments.Its target artifacts include web-grounded answers, reports, slides, memos, structured documents, financial workbooks, and presentations.
  • 4.2.1. Deep Search and Open-Web Research: Deep-search tasks scale from retrieval to multi-step browsing and cross-source corroboration through guide-and-rewrite synthesis with explicit evidence specifications.Iterative rewriting controls difficulty while evidence specifications discourage plausible fabrication.
  • 4.2.2. Knowledge-Worker Office Tasks: Office-task data extends GDPval with hierarchical occupational synthesis and rubric-based acceptance covering behaviors, errors, regional appropriateness, reasoning depth, and factuality.The corpus combines a filtered executable seed subset with broader self-synthesized coverage across industries, regions, and cultural contexts.
  • 4.2.3. Financial Analysis and Spreadsheet Operations: Financial and spreadsheet tasks are grounded in executable tool traces or workbook walks, with answers derived from observable outputs and acceptance favoring deterministic value-level matches.Coverage includes external-data retrieval and reasoning, spreadsheet manipulation, financial modeling, and workbook reconstruction.
  • 4.2.4. Slide Generation and Editing: Slide data supports both end-to-end deck creation and incremental editing through parallel synthesis streams varying source domains, request granularity, length, and language register.One stream treats slide authoring as open-ended generation, while the other addresses incremental slide editing.
  • 4.3. Reasoning-Intensive Tasks: Reasoning-intensive data scales unique problems, multiple correct solution paths, and training mixtures, with response diversity primarily improving out-of-domain generalization.Query expansion increases coverage and breadth, whereas response expansion deepens solution-strategy diversity; allocation is calibrated to capability bottlenecks.
  • 4.4. General-Purpose Conversation and Writing: General-purpose conversation and writing data combines long-CoT samples, stylistic writing, file-system interaction, tool-augmented and tool-free examples, and rigorous verification.The track aims to preserve general-purpose competence while providing a cold-start foundation for reinforcement learning and flexible interaction.
  • 4.5. Role-Play and Persona Coherence: Role-play is formalized as long-horizon conditional generation over Worlds × Stories conditioned on User Preferences, requiring physical, narrative, and stylistic coherence.Role-Play Bench evaluates multi-turn self-play trajectories by penalizing objectively detectable failure modes because alignment is subjective.

5. Supervised Fine-Tuning

The SFT stage instills interleaved thinking in M2 as a foundation for subsequent RL. Its data pipeline produces high-quality trajectories by combining domain-specific rejection sampling with multi-stage cleaning across chat, reasoning, code, and cowork.

  • Supervised Fine-Tuning: SFT instills desired interleaved thinking behavior in M2 and provides a strong starting point for subsequent RL.The training objective differs from conventional long CoT data by integrating reasoning into broader trajectories.
  • Supervised Fine-Tuning: Interleaved SFT trajectories combine thinking traces with intermediate actions and observations, enabling the model to reason, act, and revise.This unified trajectory structure contrasts with reasoning confined to a single contiguous block.
  • Supervised Fine-Tuning: Large-scale rejection sampling against domain-specific reward signals followed by multi-stage data cleaning yields high-quality interleaved-thinking trajectories.The pipeline is designed to produce such data at scale.
  • Supervised Fine-Tuning: The resulting SFT corpus spans chat, reasoning, code, and cowork, covering both single-turn reasoning.These four domains define the core coverage of the corpus.

6. Reinforcement Learning · 6.1. RL Algorithm

The RL algorithm models the LLM as a policy operating within an agent environment that includes context, memory, tools, and state transitions. It combines CISPO optimization, composite rewards, and mixed-domain training to improve long-horizon credit assignment, execution efficiency, and capability preservation.

  • 6.1.1. Agent RL Modeling: The LLM is treated as the policy, while context management, memory access, and agent state transitions constitute the environment.This abstraction extends standard reinforcement learning to agentic systems.
  • 6.1.2. MDP Formulation: Agent-environment interaction is formulated as an MDP in which each single-step LLM completion may reason, invoke tools, manage context, communicate with sub-agents, or combine these operations.The environment executes requested operations and returns observations, while the trajectory forms a complete episode parameterized by the LLM weights.
  • 6.1.2. MDP Formulation: The environment boundary is placed at the model’s generation interface, encompassing external tool execution and agent-harness control flow such as context management, branching, and delegation.This treats all components responding to model outputs as environment dynamics.
  • 6.1.3. Training Objective: Training uses individual (s_t, a_t) pairs as atomic policy-gradient samples, decoupling the policy from whether state evolution involved appending messages, truncating context, or rewriting history.The model need not explicitly control or understand the mechanics of state transitions.
  • 6.1.4. Policy Optimization: CISPO adapts clipped importance sampling with asymmetric clipping and stop-gradient weighting to produce stable first-order policy updates.The upper bound limits excessively large updates, the zero lower bound permits aggressive down-weighting, and stop-gradient prevents second-order terms.
  • 6.1.5. Reward Design: Agent trajectories may span up to 192K tokens with thousands of intermediate actions, motivating composite rewards that combine dense process supervision, completion-time incentives, and reward-to-go baselines.The completion-time reward encourages parallelism, while trajectory-level baselines improve credit assignment and stabilize optimization.
  • 6.1.6. Mixed-Domain RL Training: Mixed-domain RL jointly samples reasoning, coding, agent, and general data within each stage to avoid catastrophic forgetting and negative transfer while improving foundational and task-specific capabilities.Domain ratios and context length are adjusted across stages, with later training emphasizing agent and coding tasks; the strategy also improves end-to-end user experience.

6.2. RL Infrastructure

Forge addresses the infrastructure “impossible triangle” of throughput, training stability, and agent flexibility through decoupled modules, unified white- and black-box agent support, and scheduling and computation optimizations. Windowed FIFO balances rollout distribution consistency against throughput, while prefix tree merging accelerates training without approximation error.

  • Infrastructure objectives: The infrastructure balances system throughput, training stability, and agent flexibility, whose pairwise tensions arise from rollout-time variance, state coupling, and long-horizon credit assignment.Agent rollouts range from seconds to hours, while long-horizon trajectories reach up to 192K tokens.
  • Forge architecture: Forge decouples agent execution from training and inference through a middleware layer, allowing arbitrary agents and pipelines to scale independently.The Agent Side produces trajectories, while the Gateway Server and Data Pool bridge agents with the Rollout and Train Engines.
  • Agent flexibility: Forge’s Gateway abstraction supports both white-box agents with framework-visible context management and black-box agents exposing only observed request trajectories.This unified design was validated across hundreds of agent scaffolds and thousands of tool invocation formats.
  • Windowed FIFO: Windowed FIFO greedily fetches completed trajectories within a sliding window while enforcing ordering across window boundaries, mitigating head-of-line blocking and distribution shift.The window size W controls the trade-off between strict FIFO consistency and greedy throughput.
  • Windowed FIFO: W = 0.3N maintains near-FIFO distributional properties while substantially reducing cluster idle time.Smaller windows approach strict FIFO, whereas larger windows approach greedy scheduling.
  • Prefix tree merging: 40× training speedup is achieved by prefix tree merging, which computes shared prefixes once and preserves independent per-sample loss computation with zero approximation error.The method reduces memory consumption and enables longer sequences and larger batch sizes without compromising training fidelity.

7. Agentic Mechanism

MiniMax-M2 uses interleaved thinking with persistent reasoning state to support adaptive, self-correcting multi-step tool use. The M2.7 checkpoint extends this agentic design toward self-evolution by autonomously debugging training runs and improving its supporting scaffold.

  • Interleaved thinking: Interleaved thinking alternates reasoning tokens and action tokens, conditioning each reasoning segment on the complete prior interaction history.This lets the model revise plans, update hypotheses, and incorporate tool observations before selecting the next action.
  • Reasoning state persistence: Reasoning state persistence appends complete model outputs, including thinking blocks, to subsequent message histories instead of forcing context re-derivation.Dropping prior thinking causes cumulative state drift and degraded self-correction.
  • The Plan-Act-Reflect loop: The Plan-Act-Reflect loop makes each turn review accumulated state, execute a grounded tool call, and evaluate observations to revise or continue the plan.This operationalizes self-correction through iterative planning and reflection.
  • Effect of reasoning state persistence: Consistent gains across agentic benchmarks, especially deep search and software engineering, show reasoning-state persistence matters most for sustained multi-step planning and refinement.The ablation strips thinking blocks from prior turns before each model invocation.
  • Self-evolution: M2.7’s self-evolution shift culminates the M2 series’ progression from routine debugging and reporting to an integrated pipeline driving its own iterative development.The Model Iteration System uses human-directed goals and review while an internally generated Agent Harness supplies action chaining, persistent memory, and safety guardrails.
  • Self-evolution: 30% to 50% of daily iteration workload is absorbed as M2.7 profiles runs, diagnoses metric anomalies, debugs code, and adjusts configurations during autonomous execution.In a separate 100-round autonomous scaffold cycle, M2.7 analyzed failures, modified code, evaluated changes, and achieved a 30% performance gain on in-house evaluations.

8. Evaluation

MiniMax-M2.7 is evaluated against leading closed-weight models and MiniMax-M2.5 across agentic coding, cowork, and reasoning benchmarks, emphasizing long-horizon, environment-grounded behavior. It shows broad competitiveness, substantial gains over M2.5, and evidence of autonomous ML-engineering behavior.

  • Evaluation setup: M2.7 is compared with Claude Opus 4.6, Claude Sonnet 4.6, GPT 5.4, Gemini 3.1 Pro, and M2.5 across agentic coding, agentic cowork, and reasoning and knowledge.Closed-weight baselines use their strongest reasoning configurations, while M2.7 and M2.5 use thinking-enabled interleaved-thinking trajectories.
  • Software engineering and coding agent: 56.2 on SWE-bench Pro, 76.5 on SWE-bench Multilingual, 52.7 on Multi-SWE-bench, and 57.0 on Terminal-Bench 2.0 demonstrate broad coding-agent competitiveness.M2.7 also reaches 39.8 on NL2Repo and a 66.6 % medal rate on MLE Bench Lite.
  • Application development: 67.6 on HyperTask improves by 8 points over M2.5, while M2.7 reaches 55.6 on VIBE-Pro, on par with leading closed-weight baselines.The passage attributes this domain’s performance to approximately 10 B activated parameters, AppDev-targeted data, and Agent-as-a-Verifier reward.
  • Cowork — search and deep research: 77.8 on BrowseComp, 75.2 on Wide Search, and 64.3 on RISE show strong search performance, with RISE improving 14 points from M2.5’s 50.2.RISE requires multi-step browsing and cross-source corroboration with a Playwright browser tool.
  • Cowork — agent, office, and workspace: 63.3 on MEWC v2, 57.0 on Finance Modeling Pro, 62.7 on MM Claw, 50.0 on GDPval-AA, and 46.3 on Toolathlon mark substantial cowork capability gains.M2.5→M2.7 gains include +13.5 on MEWC v2, +23.2 on Finance Modeling Pro, and +15.0 on GDPval-AA.
  • Reasoning and self-evolution: 94.2 on AIME 2026, 89.8 on GPQA-Diamond, 76.0 on IFBench, and 72.0 on AA-LCR place M2.7 in the frontier band across reasoning-heavy benchmarks.On MLE Bench Lite, M2.7 averages a 66.6% medal rate, ties Gemini 3.1 Pro, and can debug its training scaffold, modify configuration files, and iterate over hundreds of rounds.

9. Conclusion

The MiniMax-M2 series argues that mini activations can unlock maximum real-world intelligence, exemplified by M2’s 9.8B-activated / 229.9B-total backbone. Its co-evolving data, RL, and self-evolution components translate a ∼10 B activated-parameter footprint into parity with much larger frontier systems.

  • Model thesis and architecture: M2 pairs 9.8B activated parameters with 229.9B total parameters in a Mixture-of-Experts backbone.The series is built around the thesis that mini activations can unleash maximum real-world intelligence.
  • Training system: Three co-evolving components span M2 through M2.7: agent-driven data pipelines, Forge’s agent-native RL system, and self-evolution.The data pipelines ground training trajectories in executable workspaces and artifact-aligned rewards, while Forge scales long-horizon agent trajectories.
  • Results and outlook: A ∼10 B activated-parameter footprint reaches parity with frontier systems an order of magnitude larger in per-step compute across agentic coding, agentic cowork, and reasoning & knowledge benchmarks.The paper frames these results as one step in a longer trajectory, with data, RL system, and self-evolution remaining far from saturation.

A. Contributors

The report presents its contributors in alphabetical order, with the supplied list spanning names from Aili Chen through Yuxuan Sh.

  • A. Contributors: Contributors are listed alphabetically, beginning with Aili Chen and continuing through Haonan L.The supplied passage contains the alphabetical sequence from Aili Chen to Haonan L.
  • A. Contributors: The alphabetical list continues from Han through Yuxuan Sh, including the remaining supplied contributor names.This continuation includes names such as Han, Xinyu Guan, and Yuxuan Sh.
Loading 2605.26494v2…