pith. sign in

arxiv: 2605.26494 · v1 · pith:DWUOSFPWnew · submitted 2026-05-26 · 💻 cs.AI · cs.CL· cs.LG

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

MiniMax: Aili Chen , Aonian Li , Baichuan Zhou , Bangwei Gong , Binyang Jiang , Boji Dan , Changqing Yu , Chao Wang
show 197 more authors
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 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 Lingye Bao Lingyu Yang Lingyuan Zhou Loki Lu Chen Lunbin Ceng 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 Qidi Xu Qihan Ren Qile Li Qin Wang Quanliang Chen Qunhong Ceng Rong Tian Rui Dong Ruitao Leng Ruize Zhang Shanqi Liu Shaoyu Chen Sheng Jia Shun Yao Shuoran Zhao Shuqi Yu Sichen Li Sicheng Pan Songquan Zhu Tengfei Li Tian Xie Tiancheng Qin Tianrun Liang Wei Liu Weiqi Xu Weitao Li Weixiang Chen Weiyu Cheng Weiyu Zhang Wenhu Chen Wenqian Zhao Xiancai Chen Xiangjun Song Xiangyuan Wang Xiao Luo Xiao Su Xiaobo Li Xiaodong Han Xiaojie Wu Xihao Song Xingyi Han Xinyu Guan Xuan Lu Xun Zou Xunhao Lai Xutong Li Yan Gong Yang Wang Yang Xu Yangsen Wang Ye Tang Yicheng Chen Yinran Qiu Yiqi Shi Yiting Guo Yiwen Huang Yixuan Wang Yongyi Hu Yu Gao Yu Zhang 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
This is my paper

Pith reviewed 2026-06-29 18:40 UTC · model grok-4.3

classification 💻 cs.AI cs.CLcs.LG
keywords mixture of expertsagentic language modelsmini activationsreinforcement learningself-evolutionagent trajectories
0
0 comments X

The pith

Mini activations in large MoE models reach frontier performance on agentic tasks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces the MiniMax-M2 series of Mixture-of-Experts models that keep only a small number of parameters active per token. It argues that this approach, paired with agent-driven data collection and a specialized reinforcement learning system, produces competitive results on coding, search, office, and reasoning benchmarks. The central mechanism is a pipeline that generates verifiable agent trajectories inside executable workspaces, combined with training-inference decoupling and early self-evolution steps in later checkpoints.

Core claim

The M2 series shows that a 229.9 billion parameter model activating only 9.8 billion parameters per token can reach frontier-tier results on agentic coding, deep search, office-task, and reasoning benchmarks when trained with large-scale verifiable trajectories, the Forge RL system, windowed scheduling, and initial autonomous debugging capabilities.

What carries the argument

Agent-driven data pipelines that produce verifiable trajectories in executable workspaces, paired with the Forge RL system for long-horizon adaptation and prefix-tree merging.

If this is right

  • Agent performance becomes less dependent on total parameter count and more dependent on trajectory quality and RL scheduling.
  • Training and inference can be decoupled to support both white-box and black-box agents in the same framework.
  • Later checkpoints can modify their own scaffolds through autonomous debugging of training runs.
  • Windowed-FIFO scheduling and prefix-tree merging become practical for scaling long-horizon agent training.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Sparsity in activation could become a primary lever for deploying capable agents under fixed compute budgets.
  • Focus may shift from scaling total parameters to engineering verifiable agent environments and reward signals.
  • Self-evolution loops could reduce the need for repeated human intervention in model improvement cycles.

Load-bearing premise

Agent-driven pipelines can generate large-scale, verifiable trajectories that are grounded in executable workspaces and aligned with artifact rewards.

What would settle it

An independent evaluation in which the M2 series falls below frontier models on the same agentic coding and reasoning benchmarks despite the reported activation size.

read the original abstract

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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript introduces the MiniMax-M2 series of Mixture-of-Experts language models (flagship M2: 229.9B total parameters, 9.8B activated per token) built around mini activations for agentic deployment. It rests on three components—agent-driven data pipelines for verifiable trajectories in coding and cowork tasks, the Forge RL system with windowed-FIFO scheduling and prefix-tree merging, and an early self-evolution step in M2.7—and claims this combination yields frontier-tier performance on agentic coding, deep search, office-task, and reasoning benchmarks.

Significance. If the performance claims held with supporting evidence, the work would have substantial significance for efficient agentic AI, demonstrating that sparse activation footprints can deliver high real-world capability via specialized data, RL, and self-improvement pipelines. This could influence scalable training paradigms for long-horizon agents. However, the manuscript provides no quantitative results, making any assessment of significance speculative at present.

major comments (2)
  1. Abstract: The assertion of 'frontier-tier performance' across M2 through M2.7 on agentic coding, deep search, office-task, and reasoning benchmarks is made with zero supporting numbers, tables, baselines, error bars, or methodological details, rendering the central claim load-bearing yet entirely unsupported.
  2. Abstract: The three listed components (agent-driven data pipelines, Forge RL with windowed-FIFO and prefix-tree merging, and self-evolution) are described at a high level with no equations, implementation specifics, ablation studies, or training details, preventing evaluation of how they purportedly translate mini activations into the claimed performance.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their review. We agree that the submitted manuscript lacks the quantitative results, tables, and implementation details needed to substantiate the claims, and we will revise accordingly.

read point-by-point responses

  1. Referee: [—] Abstract: The assertion of 'frontier-tier performance' across M2 through M2.7 on agentic coding, deep search, office-task, and reasoning benchmarks is made with zero supporting numbers, tables, baselines, error bars, or methodological details, rendering the central claim load-bearing yet entirely unsupported.

    Authors: We acknowledge that the current manuscript provides no numerical results, tables, or supporting evidence for the performance claims. In the revised version we will add comprehensive benchmark tables with specific scores, baselines, error bars, and methodological details for the M2 series on agentic coding, deep search, office-task, and reasoning tasks. revision: yes

  2. Referee: [—] Abstract: The three listed components (agent-driven data pipelines, Forge RL with windowed-FIFO and prefix-tree merging, and self-evolution) are described at a high level with no equations, implementation specifics, ablation studies, or training details, preventing evaluation of how they purportedly translate mini activations into the claimed performance.

    Authors: We agree the descriptions are high-level. The revision will include equations for the RL components, implementation specifics for the agent-driven data pipelines and Forge system (including windowed-FIFO scheduling and prefix-tree merging), ablation studies, and training details to enable evaluation. revision: yes

Circularity Check

0 steps flagged

No circularity: no derivations, equations, or load-bearing self-citations present

full rationale

The provided abstract and description contain no mathematical derivations, first-principles results, fitted parameters presented as predictions, or self-citations. The central claim attributes performance to three engineering components (agent-driven pipelines, Forge RL system, and self-evolution checkpoint) without any reduction to inputs by construction or renaming of known results. This is a standard empirical model announcement with no derivation chain to inspect for circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, preventing identification of specific free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 6552 in / 1017 out tokens · 37323 ms · 2026-06-29T18:40:28.654041+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 6 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. CLI-Universe: Towards Verifiable Task Synthesis Engine for Terminal Agents

    cs.AI 2026-06 unverdicted novelty 7.0

    CLI-Universe synthesizes a verified 6K dataset of terminal-agent tasks that, when used to fine-tune Qwen3-32B, reaches 33.4% on Terminal-Bench 2.0 and sets a new open-source SOTA for models at or below 32B parameters.

  2. OCELOT: Inference-Leakage Budgets for Privacy-Preserving LLM Agents

    cs.CR 2026-06 unverdicted novelty 7.0

    OCELOT recasts agent privacy as posterior-risk control and implements Witness-Verified Declassification to authorize the least-disclosing useful release under a sink-trust-weighted min-entropy budget.

  3. AutoMedBench: Towards Medical AutoResearch with Agentic AI Models

    cs.AI 2026-06 conditional novelty 7.0

    AutoMedBench evaluates AI agents on long-horizon medical workflows across five stages and finds validation and submission as dominant failure points based on thousands of runs.

  4. INCARBench: A Benchmark for Scientific Configuration in VASP INCAR by Large Language Models

    cond-mat.mtrl-sci 2026-06 unverdicted novelty 6.0

    INCARBench evaluates 19 LLMs on VASP INCAR configuration generation and repair, showing high semantic accuracy but lower scientific correctness especially for DFT+U, magnetism, and correlated materials.

  5. Vortex: Efficient and Programmable Sparse Attention Serving for AI Agents

    cs.AI 2026-06 unverdicted novelty 6.0

    Vortex provides a programmable frontend and backend for sparse attention in LLM serving, delivering up to 3.46x throughput over full attention while preserving accuracy.

  6. Token-Operations-Oriented Inference Optimization Techniques for Large Models

    cs.SE 2026-06 unverdicted novelty 3.0

    The paper introduces a four-layer technical architecture for token-operations-oriented inference optimization in large models and reviews key technologies and industry status at each layer.

Reference graph

Works this paper leans on

6 extracted references · 2 canonical work pages · cited by 6 Pith papers

  1. [1]

    GQA: Training generalized multi-query transformer models from multi-head checkpoints

    Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. GQA: Training generalized multi-query transformer models from multi-head checkpoints. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,

    2023

  2. [2]

    Dive: Scaling diversity in agentic task synthesis for generalizable tool use,

    Aili Chen, Chi Zhang, Junteng Liu, Jiangjie Chen, Chengyu Du, Yunji Li, Ming Zhong, Qin Wang, ZhengmaoZhu,JiayuanSong,etal. Dive: Scalingdiversityinagentictasksynthesisforgeneralizable tool use.arXiv preprint arXiv:2603.11076,

    work page arXiv

  3. [3]

    Webexplorer: Explore and evolve for training long-horizon web agents

    Hongru Liu et al. Wide Search: Benchmarking long-horizon multi-source web research, 2025a. Junteng Liu, Yunji Li, Chi Zhang, Jingyang Li, Aili Chen, Ke Ji, Weiyu Cheng, Zijia Wu, Chengyu Du, Qidi Xu, et al. Webexplorer: Explore and evolve for training long-horizon web agents.arXiv preprint arXiv:2509.06501, 2025b. Xianzhen Luo, Jingyuan Zhang, Shiqi Zhou,...

    work page arXiv

  4. [4]

    American invitational mathematics examination 2025.https: //www.maa.org/math-competitions/aime,

    Mathematical Association of America. American invitational mathematics examination 2025.https: //www.maa.org/math-competitions/aime,

    2025

  5. [5]

    American invitational mathematics examination 2026.https: //www.maa.org/math-competitions/aime,

    Mathematical Association of America. American invitational mathematics examination 2026.https: //www.maa.org/math-competitions/aime,

    2026

  6. [6]

    Le, Ed H

    Mirac Suzgun, Nathan Scales, Nathanael Schärli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V. Le, Ed H. Chi, Denny Zhou, and Jason Wei. Challenging BIG-Bench tasks and whether chain-of-thought can solve them. InFindings of the Association for Computational Linguistics: ACL 2023,

    2023