The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence
by MiniMax
The full training account of a 229.9B-parameter MoE with 9.8B active — built end-to-end for agentic work.
Overview
Nine sections, and the interesting material is in the back half. Section 2 covers a Mixture-of-Experts architecture with 229.9B total parameters and 9.8B activated per token, 256 fine-grained experts with sigmoid gating, 48 query heads and 8 key-value heads under GQA, and multi-token prediction of three future tokens for speculative decoding; notably the team kept full attention on every layer after finding sliding-window and sparse variants fragile in production agent loops. Section 3 describes pre-training on 29.2T tokens with progressive context extension to 192K. Section 4 details four agentic data pipelines: SWE tasks built from real GitHub pull requests into verifiable fail-to-pass and pass-to-pass repair tasks, full-stack AppDev tasks scored by an Agent-as-a-Verifier across execution, interaction and aesthetics, a synthesized Terminal-Gym, and agentic cowork spanning deep search, GDPval-grounded office tasks, financial analysis and slide generation. Section 6 documents Forge, the RL system: a CISPO optimizer with asymmetric clipping, composite process, completion-time and performance rewards, a decoupled agent-middleware-trainer architecture supporting both white-box and black-box agents, windowed-FIFO scheduling and prefix-tree merging reported at up to 40x speedup. Section 7 covers interleaved thinking and the M2.7 self-evolution loop, which the authors say absorbed 30-50% of daily iteration work. Section 8 benchmarks against Claude Opus 4.6, GPT 5.4 and Gemini 3.1 Pro. Submitted May 2026, revised July 2026.
At a Glance
- Topic
- Models
- Level
- Advanced
- Format
- Paper
- Cost
- Free
- Duration
- ~3-4 hours for a careful full read; ~45 minutes for the RL and evaluation sections alone
- Provider
- MiniMax
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Why MiniMax rejected sliding-window and sparse attention for production agent loops
- ✓How to convert real GitHub pull requests into verifiable fail-to-pass training tasks
- ✓The CISPO policy-optimization objective and its asymmetric importance-sampling clipping
- ✓How composite rewards mix process signals, completion time and task performance
- ✓Decoupling agent, middleware and trainer so black-box scaffolds can still be RL-trained
- ✓Windowed-FIFO scheduling and prefix-tree merging as agent-RL throughput techniques
- ✓How interleaved thinking preserves reasoning state across multi-turn tool calls
- ✓Where a 9.8B-active open-weight model lands against Opus 4.6, GPT 5.4 and Gemini 3.1 Pro
Highlights
- •One of the few 2026 reports that documents the RL infrastructure and not just the algorithm — scheduling, prefix merging and training/inference decoupling are all specified
- •Reports a negative result worth knowing: efficient attention variants proved fragile when KV-like state is held in lower precision inside long agent loops
- •The self-evolution section, in which M2.7 autonomously debugs training runs and rewrites its own scaffold for a reported 30% internal gain, has no real equivalent in other open reports
- •The weights are public and heavily used — MiniMax-M2.7 on Hugging Face shows roughly 920K downloads in a month — so the claims are independently checkable
- •Full within-series progression from M2 to M2.5 to M2.7 lets you see what each post-training change actually bought
Who It’s For
Best For
- ✓Researchers and engineers designing an agent-RL or post-training stack
- ✓Teams evaluating open-weight models for agentic coding workloads
- ✓Infrastructure engineers needing throughput patterns for long-horizon RL rollouts
- ✓Anyone weighing sparse versus full attention for tool-using agents
Prerequisites
- •Comfortable reading ML papers with RL notation and Mixture-of-Experts terminology
- •Familiarity with post-training: SFT, policy-gradient RL and reward modelling
- •Helpful: prior exposure to SWE-bench-style agent evaluation
FAQ
What is The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence?
The technical report behind one of the most widely used open-weight agentic model families of 2026, covering the flagship M2 through the M2.7 checkpoint. It is written for engineers and researchers who want to know how a frontier-tier agentic coding model is actually built — data pipelines, RL infrastructure, reward design and scheduling — rather than what it scores. Read it if you are designing your own post-training or agent-RL stack.
Is The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence free?
The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence is free to access.
What level is The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence for?
The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence is aimed at a advanced audience. Recommended background: Comfortable reading ML papers with RL notation and Mixture-of-Experts terminology, Familiarity with post-training: SFT, policy-gradient RL and reward modelling, Helpful: prior exposure to SWE-bench-style agent evaluation.
How long does The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence take?
Expect roughly ~3-4 hours for a careful full read; ~45 minutes for the RL and evaluation sections alone. Most learners work through it at their own pace.
What will I learn from The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence?
You'll learn: Why MiniMax rejected sliding-window and sparse attention for production agent loops; How to convert real GitHub pull requests into verifiable fail-to-pass training tasks; The CISPO policy-optimization objective and its asymmetric importance-sampling clipping; How composite rewards mix process signals, completion time and task performance; Decoupling agent, middleware and trainer so black-box scaffolds can still be RL-trained; Windowed-FIFO scheduling and prefix-tree merging as agent-RL throughput techniques; How interleaved thinking preserves reasoning state across multi-turn tool calls; Where a 9.8B-active open-weight model lands against Opus 4.6, GPT 5.4 and Gemini 3.1 Pro.
Topics
Sources
This page was written from 3 sources, 1 on domains other than arxiv.org.