DEV Community

Cover image for MiniMax M2.7 Is Live on BlockRun — The First Self-Evolving Reasoning Model
1bcMax
1bcMax

Posted on • Originally published at blockrun.ai

MiniMax M2.7 Is Live on BlockRun — The First Self-Evolving Reasoning Model

#ai #llm #machinelearning #openai

MiniMax just dropped M2.7 — and it's live on BlockRun right now.

One API call. Pay per request. No subscription. No API key signup with MiniMax.

curl https://blockrun.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax/minimax-m2.7",
    "messages": [{"role": "user", "content": "Hello"}]
  }'
Enter fullscreen mode Exit fullscreen mode

If you're still calling minimax/minimax-m2.5, it auto-redirects to M2.7. No code changes needed.

What Makes M2.7 Different

M2.7 is the first model MiniMax describes as deeply participating in its own evolution. It doesn't just run agent tasks — it builds and optimizes its own agent harnesses through recursive self-improvement loops.

In practice, that means:

  • 97% skill adherence across 40+ complex skills (each exceeding 2,000 tokens)
  • 30% performance gains from recursive harness optimization over 100+ iteration cycles
  • Handles 30–50% of research workflows autonomously

This isn't a chatbot upgrade. It's a model that gets better at being an agent the more you use it as one.

Benchmarks That Matter

MiniMax M2.7 benchmarks vs Sonnet 4.6, Opus 4.6, Gemini 3.1 Pro, GPT-5.4 across SWE Bench Pro, Multi-SWE Bench, VIBE-Pro, MLE-Bench Lite, GDPval-AA, Toolathon, MM-ClawBench, and Artificial Analysis

Software Engineering

Benchmark M2.7 Context
SWE-Pro 56.22% Matches GPT-5.3-Codex
VIBE-Pro 55.6% End-to-end project delivery
Terminal Bench 2 57.0% Complex engineering systems
SWE Multilingual 76.5 Cross-language code tasks
Multi SWE Bench 52.7 Multi-repo engineering

Machine Learning & Research

MLE Bench Lite (22 Kaggle-style competitions): 66.6% average medal rate — second only to Opus 4.6 (75.7%) and GPT-5.4 (71.2%). Best single run: 9 gold, 5 silver, 1 bronze.

Professional Productivity

Benchmark M2.7 Context
GDPval-AA ELO 1495 Highest among open-source models
Toolathon 46.3% Tool use accuracy
MM Claw 62.7% Near Sonnet 4.6 level

Production debugging benchmarks show incident recovery time under 3 minutes — SRE-level decision-making for log analysis, security audits, and system comprehension.

New in M2.7 vs M2.5

  • Native Agent Teams — multi-agent collaboration built into the model, not bolted on
  • Recursive self-improvement — the model optimizes its own harnesses over iteration cycles
  • Character consistency — dramatically improved emotional intelligence for interactive apps
  • Financial analysis — deep reasoning over complex financial documents and reports

Pricing on BlockRun

Price
Input $0.30 / 1M tokens
Output $1.20 / 1M tokens
Context window 204,800 tokens

That's 50x cheaper than Claude Opus and 12x cheaper than GPT-5.4 for output tokens — while matching their engineering benchmarks.

Pay per request with USDC on Base. No API key. No subscription. No minimum spend.

Try It Now

Direct API:

https://blockrun.ai/v1/chat/completions
Enter fullscreen mode Exit fullscreen mode

Python SDK:

from blockrun import BlockRun
client = BlockRun()
response = client.chat("minimax/minimax-m2.7", "Explain this codebase")
Enter fullscreen mode Exit fullscreen mode

TypeScript SDK:

import { BlockRun } from "blockrun";
const client = new BlockRun();
const response = await client.chat("minimax/minimax-m2.7", "Explain this codebase");
Enter fullscreen mode Exit fullscreen mode

ClawRouter (drop-in OpenAI replacement for any framework):

export OPENAI_BASE_URL=https://blockrun.ai/v1
# Works with OpenClaw, LangChain, CrewAI, AutoGen — any OpenAI-compatible client
Enter fullscreen mode Exit fullscreen mode

Read the full announcement from MiniMax: MiniMax M2.7 — Beginning the Journey of Recursive Self-Improvement

Top comments (0)