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. API Yi 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
| Benchmark | M2.7 Score | Notes |
|---|---|---|
| SWE-bench Pro | 56.22% | Near Opus-level |
| SWE-bench Verified | 78% | Strong software engineering |
| VIBE-Pro | 55.6% | End-to-end project delivery |
| Terminal Bench 2 | 57.0% | Complex engineering systems |
| MM Claw | 62.7% | Agent tasks |
| MLE Bench Lite | 66.6% | ML competition medal rate |
| Skill Adherence | 97% | Across 40+ complex skills |
M2.7 vs M2.7-highspeed
Both variants produce identical output quality — the difference is speed and cost:| Aspect | M2.7 Standard | M2.7-highspeed |
|---|---|---|
| Output Quality | Identical | Identical |
| Speed | ~60 TPS | ~100 TPS |
| Context Window | 204,800 tokens | 204,800 tokens |
| Best For | Budget-conscious | Latency-sensitive production |
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
| Model | Input Price | Output Price | Billing Type |
|---|---|---|---|
| MiniMax-M2.7 | $0.30 / 1M tokens | $1.20 / 1M tokens | Pay-per-token - Chat |
| MiniMax-M2.7-highspeed | $0.60 / 1M tokens | $2.40 / 1M tokens | Pay-per-token - Chat |
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