MiniMax Launches M2.7, Its First Self-Evolving Agentic Model

MiniMax

Mar 19, 2026 · Updated Apr 25, 2026

MiniMax released M2.7, their first model that actively participated in its own training evolution. It achieves SOTA on SWE-Pro (56.2%) and matches Claude Sonnet 4.6 in OpenClaw for agentic coding tasks, with API and MiniMax Agent access available now.

MiniMax-M2.7, MiniMax's latest model, is the first in their lineup built with an autonomous internal harness that collects feedback, builds evaluation sets, and iterates on its own architecture, skills, and memory mechanisms. On benchmarks, M2.7 scores 56.2% on SWE-Pro, 57% on Terminal Bench 2, and 62.7% on MM-ClawBench — with an 88% win-rate against M2.5 head-to-head. It achieves 97% skill adherence across 40+ complex skills in agent team settings.

This positions M2.7 as a direct competitor to frontier coding and agentic models. The model matches Claude Sonnet 4.6 in OpenClaw — the primary agentic coding benchmark cited. The self-evolving harness approach signals a shift — instead of purely human-curated training, models now help define the tasks they're evaluated and improved on.

Try M2.7 through the MiniMax Agent platform, which powers the Air, Max, and MaxClaw agents. API access is available for developers building their own agentic pipelines.

View the full update on agent.minimax.io

MiniMax (official)

@MiniMax_AIMar 18

Introducing MiniMax-M2.7, our first model which deeply participated in its own evolution, with an 88% win-rate vs M2.5 - Production-Ready SWE: With SOTA performance in SWE-Pro (56.22%) and Terminal Bench 2 (57.0%), M2.7 reduced intervention-to-recovery time for online incidents to 3-min on certain occasions. - Advanced Agentic Abilities: Trained for Agent Teams and tool search tool, with 97% skill adherence across 40+ complex skills. M2.7 is on par with Sonnet 4.6 in OpenClaw. - Professional Workspace: SOTA in professional knowledge, supports multi-turn, high-fidelity Office file editing. MiniMax Agent: https://t.co/aIzrFYcfUz API: https://t.co/fHRdSV7BwZ Token Plan: https://t.co/BDCycxepZw

236

View on X

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

See all AI news & updates from MiniMax →

Keep reading

MiniMax-M2.7 Now Available on Ollama Cloud for Agentic Coding

MiniMax-M2.7 is now available on Ollama's cloud for coding and agentic tasks, with SOTA scores on SWE-Pro (56.22%) and Terminal Bench 2 (57.0%). You can run it inside Claude Code or OpenClaw without a local model download.

RunwareMar 20

Runware Adds MiniMax M2.7 on Day Zero of Release

Runware, an AI model API, now hosts MiniMax M2.7 — a long-context LLM built for agentic coding, tool use, and office productivity. The model scores 56.22% on SWE-Pro and a 97% skill adherence rate across 40+ complex tasks.

MiniMax M2.7 Enters Code Arena Top 10 as Most Cost-Efficient Model

Arena.aiMar 20

MiniMax M2.7 Enters Code Arena Top 10 as Most Cost-Efficient Model

MiniMax M2.7 has entered Code Arena's Agentic WebDev leaderboard at a preliminary rank of #8 with a score of 1452, priced at $0.30/$1.20 per million tokens — the lowest cost in the top 10, over 16x cheaper than top-ranked models.

OpenHandsMay 28

OpenHands Partners With MiniMax to Offer Free Agentic Coding Access

OpenHands integrated the MiniMax-M2.7 model and is providing free access to its autonomous coding agents for a limited time. This partnership allows developers to test the performance of cost-effective, agent-native models in real-world software development workflows for free.