Code with MiniMax M2.7
& M2.7 Highspeed in AdaL CLI
The first self-evolving AI model family. MiniMax M2.7 and MiniMax M2.7 Highspeed — 56.22% SWE-Pro, 76.5% multilingual coding, 66.6% ML research medal rate. Released March 18, 2026 — early access available now.
No credit card required
Why developers choose MiniMax M2.7 & M2.7 Highspeed
Early echoes of self-evolution
Self-Evolving AI
M2.7 is the first model that deeply participates in its own improvement — building agent harnesses and recursively optimizing its own memory and skills.
Native Agent Teams
Natively supports multi-agent collaboration with role differentiation. Handles 30–50% of ML research workflows autonomously.
Production-Grade Coding
Reduced incident recovery to <3 minutes in live production debugging. Delivers full Web, Android, iOS, and simulation projects end-to-end.
97% Skill Adherence
Maintains 97% adherence rate across 40+ complex skills (each >2,000 tokens) with dynamic tool selection from complex environments.
Multilingual Coding SOTA
76.5% on SWE Multilingual — purpose-built for internationalized codebases and non-English engineering teams.
ML Research Automation
Automates literature review, experiment specification tracking, statistical trace analysis, and code merge request generation with context awareness.
MiniMax M2.7 Model Variants
Choose the right variant for your workflow
MiniMax M2.7
Maximum capability for complex agentic tasks, multi-repo coding, and ML research automation. Best for long-horizon tasks requiring deep reasoning and self-evolution.
MiniMax M2.7 Highspeed
Faster inference for interactive coding, quick edits, and high-throughput agentic loops. Same M2.7 generation with latency tuned for real-time developer workflows.
Software Engineering Performance
Approaching frontier model levels across coding benchmarks
| Benchmark | MiniMax M2.7 | Notes |
|---|---|---|
| SWE-Pro (Agentic coding) | 56.22% | Approaches Opus level |
| VIBE-Pro (End-to-end project delivery) | 55.6% | |
| Terminal Bench 2 (Complex systems) | 57.0% | |
| SWE Multilingual | 76.5% | |
| Multi SWE Bench (Real-world scenarios) | 52.7% |
Professional & Research Performance
SOTA on office work, global top tier for tool use and ML research
| Benchmark | MiniMax M2.7 | Notes |
|---|---|---|
| GDPval-AA ELO (Office work) | 1495 | Highest among open-source models |
| Toolathon (Tool use) | 46.3% | Global top tier |
| MM Claw | 62.7% | Close to Sonnet 4.6 level |
| MLE Bench Lite (avg medal rate, 3 trials) | 66.6% | 2nd only to Opus 4.6 (75.7%) |
How to use MiniMax M2.7 & M2.7 Highspeed
Get started in minutes
Install AdaL CLI
Install globally with npm. Works on macOS, Linux, and Windows.
Sign up (free)
Create your account — takes less than a minute.
Select Your Variant
Use /model to pick MiniMax M2.7 or MiniMax M2.7 Highspeed from the model selector.
Ready to code with self-evolving AI?
Be among the first to use MiniMax M2.7 and MiniMax M2.7 Highspeed — the models that improve themselves — in AdaL CLI.
Get Early Access