Key Highlights
- Smallest Tier-1 Model: Only 10B active parameters, SWE-bench Pro 56.22%, SWE-bench Verified 78%, matching Opus-level performance
- Self-Evolution: Industry’s first model deeply participating in its own training, autonomously handling 30-50% of its RL development workflow
- Native Multi-Agent: Built-in Agent Teams collaboration, 40+ complex skills with 97% adherence rate
- Extreme Value: Input $0.30 / Output $1.20 per million tokens — roughly 1/50th of comparable competitors
- Two Variants: Standard and highspeed produce identical output quality; highspeed runs at ~100 TPS
Background
On March 18, 2026, MiniMax officially released the M2.7 series. Dubbed the “smallest Tier-1 model,” M2.7 achieves performance comparable to Claude Opus 4.6 and GPT-5.3 Codex on major benchmarks with just 10B active parameters. M2.7’s standout feature is its “self-evolution” capability — it autonomously triggers log analysis, debugging, and metric evaluation, independently handling 30-50% of its own reinforcement learning development workflow, including analyzing its own failures, rewriting code segments, running evaluations, and deciding what to keep or discard. APIYI has launched both M2.7 and M2.7-highspeed with pay-per-token Chat billing.Detailed Analysis
Core Features
Self-Evolving
First model to deeply participate in its own training, autonomously handling 30-50% of RL workflow
Native Multi-Agent
Built-in Agent Teams with role boundaries, adversarial reasoning, and protocol adherence as internalized capabilities
Minimal Parameters
Just 10B active parameters achieving Tier-1 performance — extremely efficient
Advanced Tool Use
Manages 40+ complex skills (each exceeding 2,000 tokens) with 97% adherence rate
Benchmark Performance
Scores 50 on the Artificial Analysis Intelligence Index, tying with GLM-5, ahead of MiMo-V2-Pro (49) and Kimi K2.5 (47), while using 20% fewer output tokens at less than one-third the cost.
M2.7 vs M2.7-highspeed
Both variants produce identical output quality — the difference is speed and cost:Technical Specifications
- Context Window: 204,800 tokens (~205K)
- Max Output: 131,072 tokens
- Reasoning: Supports mandatory reasoning with
<think>tags - Architecture: MoE (Mixture of Experts)
Pricing & Availability
M2.7-highspeed is approximately 1.7x faster than the standard version, ideal for latency-sensitive production workloads. Both variants are identical in intelligence — choose based on your needs.
Summary & Recommendations
MiniMax-M2.7 delivers Tier-1 performance with just 10B active parameters — a remarkable achievement in efficiency. Its self-evolution capability and native multi-agent collaboration are unique differentiators, excelling in software engineering, tool calling, and complex workflow orchestration. Recommended Use Cases:- Developers needing high intelligence on a budget
- Agent workflows and multi-step task orchestration
- Software engineering assistance and code generation
- Production environments requiring strong tool-calling capabilities