New MiniMax M2.7 proprietary AI model is 'self-evolving' and can perform 30-50% of reinforcement learning research workflow

2026-03-18 · Source: VentureBeat · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Advanced, medium

Summary

MiniMax, a Chinese AI startup, has released its new proprietary large language model (LLM), M2.7, which features "self-evolving" capabilities. This model can autonomously manage 30-50% of its own reinforcement learning research workflow, including data pipelines, training environments, and evaluation infrastructure, by analyzing failure trajectories and planning code modifications over iterative loops. M2.7 is designed for AI agents and third-party tools, offering intelligence comparable to Google's Gemini 3.1 and Anthropic's Claude Opus 4.6 in autonomous research tasks, achieving a 66.6% medal rate on MLE Bench Lite. It also demonstrates significant performance gains over its predecessor, M2.5, in software engineering (56.22% on SWE-Pro), professional office tasks (Elo score 1495 on GDPval-AA), and hallucination reduction (34% rate). The model is available via API and third-party providers like OpenRouter, priced at $0.30 per 1 million input tokens and $1.20 per 1 million output tokens, making it one of the most affordable frontier AI models globally.

Key takeaway

For CTOs and VP of Engineering evaluating AI investments, MiniMax M2.7 signals a shift towards self-evolving models that can significantly reduce operational costs and accelerate development cycles. Your teams should consider integrating M2.7 for its demonstrated ability to autonomously manage research workflows and its cost-efficiency for high-level reasoning, especially in software engineering and professional office tasks. However, be mindful of its proprietary nature and Chinese origin, which may pose compliance considerations for highly regulated or government-facing industries in the U.S. and the West.

Key insights

MiniMax M2.7 introduces self-evolving AI capabilities, autonomously managing significant portions of its own development workflow.

Principles

• AI models can be architects of their own progress.
• Recursive self-improvement drives faster iteration curves.

Method

Earlier M2.7 versions built a research agent harness to manage data, training, and evaluation, autonomously triggering log-reading, debugging, and metric analysis to optimize programming performance.

In practice

• Integrate M2.7 via API for cost-efficient, high-level reasoning.
• Utilize M2.7 for professional document workflows and financial modeling.
• Explore M2.7's agentic capabilities for SRE and DevOps automation.

Topics

• Self-Evolving AI
• Large Language Models
• AI Agents
• Reinforcement Learning
• Model Benchmarking

Code references

• openai/mle-bench

Best for: CTO, VP of Engineering/Data, Machine Learning Engineer, AI Engineer, MLOps Engineer, Director of AI/ML

Related on AIssential

• M2.7 just BROKE the Entire Industry... — Miniax's M2.7 model demonstrates autonomous self-improvement, achieving significant performance gains and nearing top-tier AI models on consumer-grade hardware.
• HyEvo for Topological Reasoning Graph Optimization — A self-evolving system autonomously designs optimal hybrid LLM-code agent workflows, eliminating manual architecture design.
• MiniMax M2.7 — The Loop of Progress — MiniMax M2.7 introduces self-evolution, allowing AI to participate in its own development and accelerate progress.
• Chinese AI model MiniMax M2.7 reportedly helped develop itself — MiniMax M2.7 demonstrates AI self-improvement by autonomously optimizing its own training and development processes.
• MiniMax's new open M2.5 and M2.5 Lightning near state-of-the-art while costing 1/20th of Claude Opus 4.6 — MiniMax's M2.5 models offer near state-of-the-art AI performance at a fraction of the cost, enabling widespread agentic applications.
• MiniMax 2.7: GLM-5 at 1/3 cost SOTA Open Model — AI model development is shifting towards self-evolving agents, efficiency, and hybrid architectures.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by VentureBeat.