Shanghai-based AI lab MiniMax has officially released the open-source weights for its next-generation reasoning model, MiniMax-M2.7. Released on March 18, 2026, and broadly available through NVIDIA’s open-source inference ecosystem as of this week, M2.7 is being hailed as the first “self-evolving” model designed to actively manage its own training and engineering workflows.
Unlike previous versions, M2.7 is built on the OpenClaw (Agent Harness) framework, allowing it to function as an autonomous agent that can plan, execute, and refine complex digital tasks with minimal human intervention
1. The “Self-Evolving” Breakthrough
The defining feature of M2.7 is its ability to participate in its own evolution. During its 34-day training run, MiniMax reported that the model autonomously managed 30–50% of the routine ML engineering work.
- Autonomous Debugging: The model monitors its own training logs and hardware metrics. If it encounters a loss spike or a corrupted data batch, it attempts to diagnose and apply a fix rather than simply alerting a human engineer.
- Recursive Improvement: M2.7 can refine its own “scaffold”—the external code that helps it interact with tools—by running optimization cycles until it achieves a specific performance goal.
- Persistent Learning: Using the OpenClaw framework, the model builds reusable skills from its experiences, maintaining memory across sessions to improve its long-horizon planning.
2. Benchmark Performance: Professional Prowess
M2.7 is optimized for “production-grade” tasks rather than just chat. It uses a sparse Mixture of Experts (MoE) architecture, activating only 10 billion parameters per inference, which makes it remarkably efficient.
| Benchmark | M2.7 Score | Context / Comparison |
| GDPval-AA | 1495 ELO | Highest among open-source models for office tasks (Excel, Word). |
| SWE-Pro | 56.22% | Rivals Claude Opus 4.6 and GPT-5 on coding agents. |
| SWE-bench Verified | 78% | Significantly outperforms Claude Opus 4.6 (55%) in end-to-end patches. |
| GPQA Diamond | 87.4% | Demonstrates elite graduate-level scientific reasoning. |
| Context Window | 204,800 | Handles entire codebases and massive document sets in one session. |
3. Disruption in Speed and Pricing
By activating only a subset of its parameters, M2.7 achieves speeds and costs that challenge the current market leaders.
- Speed: Clocked at 100 Tokens Per Second (TPS), making it roughly 3x faster than Claude Opus 4.6.
- Cost Collapse: Priced at $0.30 per million input tokens and $1.20 per million output tokens, it is nearly 50x cheaper for high-volume agent workloads than the top closed-source models.
- Blended Caching: With effective cache hit rates of ~77%, the blended cost can drop as low as $0.06 per million tokens.
4. Interactive “OpenRoom” Demo
Alongside the model, MiniMax open-sourced OpenRoom, an interactive agent demonstration. Most of the code for this environment was generated by M2.7 itself. It moves AI interaction beyond text boxes into graphical environments where the agent can move, manipulate objects, and perform visual tasks, demonstrating “extreme physics mastery” and instruction following.
5. Two Model Variants
MiniMax is offering the M2.7 series in two distinct flavors:
- MiniMax-M2.7: The standard reasoning-heavy model built for professional office delivery and complex engineering.
- MiniMax-M2.7-Highspeed: Optimized for ultra-low latency, focusing on polyglot code mastery and precision refactoring for live dev environments.
“We are moving from AI that answers questions to AI that manages systems,” stated a MiniMax lead researcher. “M2.7 is designed to handle what it can autonomously and escalate to humans only what genuinely requires judgment.”
