On March 18, 2026, Chinese AI unicorn MiniMax officially launched its next‑generation Agent flagship large model MiniMax M2.7. This is the latest iteration of the M2 series and the industry's first public demonstration of a complete self‑evolution path where the model actively participates in its own training and optimization. The company emphasizes that M2.7 autonomously constructs complex Agent Harness (scaffolding) and, combined with native capabilities such as Agent Teams, multiple complex Skills, and Tool Search, can undertake 30%–50% of R&D workload in highly complex productivity tasks and achieves roughly a 30% performance improvement on internal evaluation sets. The release directly targets popular Agent frameworks like OpenClaw and MaxClaw, has been deeply integrated into the MiniMax Agent product line, and has been dubbed the "productivity nuclear weapon of the Agent era" by the developer community.
From M2.5 to M2.7: an Upgrade in MiniMax's Agent Strategy
MiniMax previously rose to prominence in Agent scenarios with M2.5: after its February 2026 release, OpenRouter call volume exceeded 3.07 trillion tokens within a week. With an extremely low running cost (just $1 per hour for continuous operation) and an SWE‑Bench Verified coding performance of 80.2%, M2.5 quickly topped global call charts. M2.5 was positioned as "the world's first production‑grade flagship model natively designed for agent scenarios," focusing on real‑world tasks like programming, office work, and multi‑turn tool invocation.
M2.7 represents a qualitative leap — no longer just an "executing Agent" but an "evolving Agent." MiniMax now uses the M2 series models internally as research‑oriented Agent frameworks: they can interact with different project teams, covering the full chain from data pipelines, training environments, and evaluation infrastructure to cross‑team collaboration and persistent memory. Researchers only need to set high‑level directions while the model autonomously delivers the next‑generation iterations. The M2.7 self‑evolution practice is presented as a public example of MiniMax's transformation into an AI‑native organization.
Core Technology: Agent Harness + Self‑Evolution Loop
The standout feature of M2.7 is its ability to autonomously construct complex Agent Harness. Traditional Agents rely on human‑designed scaffolding (Skills, memory systems, tool‑invocation flows), whereas M2.7 can generate and iterate these components on its own.
The specific mechanisms include:
- Native support for Agent Teams: the model can concurrently orchestrate multiple role personas to perform role anchoring, adversarial reasoning, protocol compliance, and differentiated behaviors, without extra prompting.
- Complex Skills and Tool Search: supports long‑form workflows over 2000+ tokens with a 97% skill compliance rate. In production incident debugging, it can autonomously perform causal reasoning, statistical analysis, database queries, root cause localization, and propose non‑blocking remediation plans, reducing recovery time to under three minutes.
- Reinforcement learning (RL) closed loop: when developing itself, M2.7 autonomously constructs dozens of complex Skills, updates the Memory system, drives reinforcement learning, and optimizes the Harness based on outcomes. Internal practices show it has completed over 100 iteration cycles: analyze failure traces → plan modifications → change code → run evaluations → decide to keep or roll back, significantly improving stability.
Official example: on MLE Bench Lite (22 machine learning tasks), a simple harness guided M2.7 through three rounds of 24‑hour self‑optimization, resulting in an average medal rate of 66.6% (9 gold, 5 silver, 1 bronze), ranking just behind Opus 4.6 and GPT‑5.4 and on par with Gemini 3.1.
Additionally, M2.7 delivers high‑fidelity multi‑turn editing in office scenarios: it can read annual reports, perform cross analyses, design hypotheses, build financial models, and generate usable PPT/Word/Excel drafts that practitioners rated as "ready for the next workflow stage."
Benchmark Results: Multi‑Dimensional Leadership, Production‑Ready
M2.7 demonstrates overwhelming advantages across multiple Agent and coding benchmarks (significant improvements over M2.5 and comparable to leading closed‑source models):
- SWE‑Pro (multilingual programming): 56.22% (tying GPT‑5.3‑Codex)
- VIBE‑Pro (repo‑level code generation): 55.6% (approaching Opus 4.6)
- Terminal Bench 2: 57.0%
- Multi SWE Bench: 52.7%
- Toolathon (tool invocation): 46.3% (world‑leading)
- MM‑Claw (OpenClaw‑compatible eval): 62.7% (near Sonnet 4.6)
- GDPval‑AA (ELO score): 1495 (highest among open models)
Across 40 complex skill cases (>2000 tokens), the skill compliance rate remains a steady 97%. These metrics indicate that M2.7 has truly evolved from a "conversational model" into a "productivity Agent."
Deployment Ecosystem: Seamless Integration with MiniMax Agent and OpenClaw
M2.7 has been fully enabled across the MiniMax Agent product range (Air, Max, MaxClaw):
- MaxClaw: one‑click cloud OpenClaw deployment, defaulting to M2.7 and offering a subscription‑based low‑cost plan.
- Skills ecosystem: deep optimizations for MaxClaw skill execution, with an official Skills Library and community contributions opening soon.
- Enterprise applications: internally covers nearly 90% of employee scenarios for programming, data analysis, and recruitment, accelerating MiniMax's own transition to an AI‑native organization.
Developers can call the API directly via platform.minimax.io, replacing existing Agent frameworks with zero friction and benefiting from the model's "self‑evolution" capabilities.
Industry Impact: A Chinese Technological Statement for the Agent Era
The M2.7 launch arrives at a pivotal moment as China’s AI call volume surpassed the United States for the first time (February saw 5.16 trillion tokens of weekly model calls in China). The MiniMax M2 series has topped OpenRouter charts for consecutive weeks, and M2.7 further consolidates its leadership in the Agent track.
Compared with competitors like Moonshot Kimi, Zhipu GLM, and DeepSeek, MiniMax’s core differentiator is "model self‑evolution" — not only a technical breakthrough but an open exploration of whether AI can autonomously iterate itself. Industry analysts believe this path could spawn the next generation of a fully automated AI R&D closed loop, greatly reducing human intervention costs.
For developers and enterprises, M2.7 means:
- Over 30% productivity gains for Agents under the same budget;
- Complex office and engineering tasks can reach "AI intern" level autonomous delivery;
- Combined with the OpenClaw ecosystem, enables low‑cost, one‑to‑one Agent deployment.
Conclusion: From Executing Agents to Evolving Agents
The significance of MiniMax M2.7 goes beyond a new model; it represents a paradigm shift in the Agent era. It demonstrates that large models are no longer passive tools but can be self‑driven, evolving productivity entities. In 2026, when compute costs remain high, this kind of intrinsic optimization capability will become one of China’s strongest competitive advantages on the global stage.
With M2.7's APIs fully open‑sourced and the Skills community launching, more developers will join this "model self‑evolution" experiment. MiniMax has again signaled through action that Chinese AI aims not only to catch up but to define the future rules of the Agent era.
▶️ Like this article? Creation takes real effort!
Subscribe to our YouTube channel:
AI Rosetta
Your likes, subscriptions and shares are our greatest support! 🙏
More AI news at: https://www.airosetta.com