MiniMax M2.7 Self-Iterates 100 Rounds, Performance Up 30%

"AI self-evolution" is a term that has been overused, but what MiniMax did with M2.7 is slightly different.

It is not about self-awareness in a sci-fi sense. Instead, M2.7, during its own training process, used an agent framework to automatically run experiments, read logs, analyze metrics, tweak code, and re-run experiments. This cycle repeated over 100 times — with no human intervention, and internal evaluation metrics ultimately improved by 30%.

Where are the boundaries of this? Is large models participating in their own iteration a direction for future AI development, or just a clever engineering maneuver?

The Specific Mechanism of "Self-Evolution"

The framework M2.7 uses is called OpenClaw. During training, a research agent harness was attached. Its job is to:

  • Monitor experiment progress
  • Automatically trigger log reading and metric analysis
  • Generate fix code after identifying issues
  • Submit fixes and re-run experiments

This loop is not designed step-by-step by humans. Instead, the agent decides the next action based on experimental results. After over 100 rounds, internal benchmarks were 30% higher than the original.

MiniMax calls this "Early Echoes of Self-Evolution" — a humble name, but the direction is significant. Engineers no longer need to watch every experiment and manually adjust parameters; the model has, to some extent, learned to improve itself.

How Does M2.7 Score?

First, programming ability, currently the most intuitive benchmark:

BenchmarkM2.7 ScoreNotes
SWE-Pro56.22%Close to Claude Opus level
VIBE-Pro (End-to-End Project Delivery)55.6%
Terminal Bench 2 (Complex System Understanding)57.0%

ML competition ability is also notable: across 22 ML competition tasks (MLE Bench Lite), the best result was 9 gold, 5 silver, and 1 bronze, with an average medal rate of 66.6% across three runs.

For general work capability, GDPval-AA ELO is 1495, currently the highest among open-source models. The execution accuracy for complex skills (each skill defined with 2000+ tokens) reaches 97%.

Independent testing by Dataconomy also shows M2.7 approaching GPT-5.3-Codex levels in software engineering tasks. Given the price difference, this comparison is quite meaningful.

Architecture: 230B Parameters, Only 10B Activated

M2.7 uses a sparse MoE (Mixture of Experts) architecture, with total parameters of 230 billion (230B) — but only 10B parameters are activated per inference.

This is the classic logic of MoE: split the model into many experts, and only call the most relevant ones for each task. The effect is that inference cost is close to a 10B dense model, while capability approaches the 230B scale.

Pricing: Input $0.30/M token, Output $1.20/M token. There is also an M2.7-highspeed version, which is faster with consistent results. Model weights were open-sourced on April 12.

Native Agent Teams: Multi-Agent Collaboration Not via Prompting

M2.7 has a specific design point: native multi-agent collaboration (Native Agent Teams).

MiniMax states that in multi-agent scenarios, some capabilities cannot be solved by prompting alone: role boundaries, adversarial reasoning, protocol adherence, behavioral differences — these must be solidified at the training level, and system prompts alone cannot support them.

This is also why M2.7 achieved 46.3% on Toolathon (tool-calling benchmark) and supports 40+ complex skills. On MiniMax's long-term product roadmap, agent capability is a core direction, and M2.7 is the furthest step taken on this path.

This Direction Deserves Serious Attention

Automated training iteration is not a new concept, but M2.7 has implemented it into a formally released commercial model, backed by concrete numbers (30% improvement, 100 rounds of iteration), rather than remaining at the paper level.

More thought-provoking is the question it opens: if models can improve their own training process, will the number of engineers at AI labs decrease in the future? MiniMax has always been a team known for small scale and high efficiency, and M2.7's automation approach is completely aligned with this DNA.

Among Chinese open-source models, M2.7 currently has the most comprehensive coverage of agentic capabilities — with data supporting multi-agent collaboration, ML competitions, and end-to-end project delivery. Unlike DeepSeek's focus on training efficiency and Qwen's focus on parameter scale, MiniMax is clearly betting that agentic workflows are the next main battlefield.

Sources: CocoLoop, MiniMax M2.7: Early Echoes of Self-Evolution (minimax.io); MiniMax M2.7 Advances Scalable Agentic Workflows on NVIDIA Platforms (NVIDIA Technical Blog); MiniMax M2.7 Matches GPT-5.3-Codex In Software Engineering Tasks (Dataconomy); MiniMax M2.7 Model Specs, Costs & Benchmarks (Galaxy.ai)