How MiniMax M2.7 Is Pioneering Self‑Evolving AI Models

01 | Milestone: AI’s First Deep Self‑Iteration

In April 2026 MiniMax open‑sourced the M2.7 model, the first self‑evolving agent that can update its own memory, acquire dozens of complex skills, and refine its learning pipeline based on experimental results.

Concrete case: an internal M2.7 version executed >100 autonomous iteration cycles while optimizing a programming scaffold, analysing failure traces, modifying code, evaluating outcomes, and deciding whether to keep or roll back changes, achieving a 30 % performance improvement without human intervention.

02 | Hard Data: What Makes M2.7 Strong?

Programming Ability – On Par with Top Closed‑Source Models

SWE‑Pro 56.22 %: matches GPT‑5.3‑Codex.

SWE‑Bench Verified 78 %: far ahead of Claude Opus 4.6 (55 %).

Terminal Bench 2 57.0 %: strong system‑level engineering understanding.

Machine‑Learning Automation – Only Behind Two Giants

On the MLE Bench Lite (22 real ML competitions) M2.7 earned a 66.6 % medal rate (9 gold, 5 silver, 1 bronze), trailing only Opus‑4.6 and GPT‑5.4.

Productivity – Best Among Open‑Source Models

GDPval‑AA ELO 1495: highest among open‑source weights, surpassing GPT‑5.3.

97 % skill compliance: stable across 40 complex >2000‑token scenarios.

Cost Advantage – 50‑60× Cheaper

Compared with Claude Opus 4.6, M2.7’s input cost is 50 × lower and output cost 60 × lower.

03 | Technical Reveal: How Self‑Evolution Works

The self‑evolution loop builds on MiniMax’s internal OpenClaw framework. The core cycle runs autonomously for >100 rounds, during which the model discovers optimisations such as:

Sampling Parameter Auto‑Tuning: systematic search for optimal temperature, frequency penalty, etc., outperforming manual tuning.

Self‑Discovered Workflows: after fixing a bug the agent automatically scans other files for the same pattern without prior instruction.

Infinite‑Loop Detection: built‑in self‑check mechanisms prevent the agent from getting stuck on complex tasks.

MiniMax estimates M2.7 handled 30‑50 % of routine ML‑engineer tasks during training, requiring researcher intervention only for critical decisions.

04 | Implications

From “Human‑Trains‑Model” to “Model‑Trains‑Model”

Current self‑evolution focuses on the agent scaffolding layer, not the model weights. If the approach scales, future versions could iterate far faster.

Qualitative Leap in Agentic Capability

M2.7 functions as an end‑to‑end engineer, delivering full project deliveries and reducing real‑world incident recovery times to under three minutes.

Open‑Source Ecosystem Impact

Despite a modified MIT license that restricts commercial use, the breakthrough has spurred rapid 0‑Day adaptations on domestic compute platforms such as Huawei Ascend and MuXi.

Technical Specs at a Glance

Architecture: 230 billion‑parameter MoE, activates 10 billion per inference.

Context window: 200 k tokens, max output 130 k tokens.

Pricing: $0.30/$1.20 per million tokens.

Open‑source repository: https://huggingface.co/MiniMaxAI/MiniMax-M2.7

Suggested inference parameters:

temperature=1.0
top_p=0.95
top_k=40

Default system prompt:

You are a helpful assistant. Your name is MiniMax-M2.7 and is built by MiniMax.