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

Hugging Face Daily Papers Papers
mixture-of-experts language-model agentic-coding reinforcement-learning self-evolution mini-activations llm-training

Summary

The MiniMax-M2 series introduces Mixture-of-Experts language models that achieve high performance on agentic tasks with minimal activated parameters (9.8B per token out of 229.9B total), leveraging agent-driven data pipelines, a scalable RL system called Forge, and a checkpoint that takes early steps toward self-evolution.

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.
Original Article
View Cached Full Text

Cached at: 05/27/26, 02:47 AM

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

Source: https://huggingface.co/papers/2605.26494 Authors:

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

Abstract

The MiniMax-M2 series introduces Mixture-of-Experts language models with minimal activated parameters that achieve high performance in agentic tasks through specialized training and deployment systems.

We introduce the MiniMax-M2 series, a family ofMixture-of-Expertslanguage 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 pipelinesproducing large-scale, verifiable trajectories acrossagentic codingandagentic cowork, each grounded in anexecutable workspaceand anartifact-aligned reward; (ii) Forge, a scalableagent-native RL systemthat adapts to long-horizon agent trajectories, paired withwindowed-FIFO scheduling,prefix-tree merging,inference optimization, and a cleantraining-inference-agent decouplingthat supports both white-box and black-box agents; (iii) the latest M2.7 checkpoint takes an early step towardself-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 onagentic coding, deep search, office-task, and reasoning benchmarks.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2605\.26494

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.26494 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.26494 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.26494 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

MiniMaxAI/MiniMax-M2.7

Hugging Face Models Trending

MiniMaxAI releases MiniMax-M2.7, an open-weight model featuring self-evolution capabilities, advanced agent team support, and strong performance on software engineering benchmarks (56.22% on SWE-Pro, 66.6% medal rate on MLE Bench Lite), with notable applications in production incident recovery and professional work tasks.

MiniMaxAI/MiniMax-M3

Hugging Face Models Trending

MiniMax releases M3, a native multimodal model with 1M context and ~428B parameters, using MiniMax Sparse Attention (MSA) for efficient long-context processing, achieving frontier-level coding and agentic performance.

MiniMax M3 (2 minute read)

TLDR AI

MiniMax introduces M3, the first open-weights model to combine coding, agentic, and multimodal capabilities with up to 1M context via sparse attention.