MiniMax-M2.7 is a large language model developed by MiniMax, noted for its deep involvement in its own evolutionary process, enabling it to handle complex agent tasks and elaborate productivity assignments. The model is capable of leveraging Agent Teams, sophisticated Skills, and dynamic tool search.

Core Methodology and Self-Evolution:
M2.7 introduces a paradigm of model self-evolution. During its development, the model was designed to autonomously update its internal memory, construct numerous complex skills for reinforcement learning (RL) experiments, and refine its own learning process based on the outcomes of these experiments. A notable instance of this self-optimization includes an internal version of M2.7 autonomously optimizing a programming scaffold over 100+ rounds. This iterative process involved the model analyzing failure trajectories, modifying its own code, running evaluations, and making autonomous decisions to either retain or revert changes, ultimately achieving a 30% performance improvement. This self-improvement capability contributes to its strong performance on the MLE Bench Lite, a benchmark comprising 22 machine learning competitions, where M2.7 achieved a 66.6% medal rate, positioning it second only to Opus-4.6 and GPT-5.4.

Professional Software Engineering Capabilities:
M2.7 demonstrates robust capabilities in real-world software engineering scenarios. These include specialized functions such as log analysis, bug troubleshooting, code refactoring, security auditing, and machine learning development. Beyond mere code generation, the model exhibits strong system-level reasoning, enabling it to correlate monitoring metrics, conduct detailed trace analyses, verify root causes in databases, and make Site Reliability Engineering (SRE)-level decisions. In practical applications, MiniMax claims M2.7 has reduced live production incident recovery time to under three minutes on multiple occasions.

Its performance on established software engineering benchmarks is competitive:

• SWE-Pro: 56.22% accuracy, matching GPT-5.3-Codex.
• Real-world engineering benchmarks: SWE Multilingual (76.5%) and Multi SWE Bench (52.7%), indicating strong performance in diverse linguistic and engineering contexts.
• VIBE-Pro: 55.6% accuracy, placing it nearly on par with Opus 4.6.
• Terminal Bench 2: 57.0%.
• NL2Repo: 39.8%.

Furthermore, M2.7 supports native Agent Teams, facilitating multi-agent collaboration with stable role identities and autonomous decision-making capabilities.

Professional Work and Tooling:
For general professional tasks, M2.7 has achieved an ELO score of 1495 on GDPval-AA, making it the highest-performing open-weight model in this category, surpassing GPT5.3. It exhibits high-fidelity multi-round editing capabilities for common office applications like Word, Excel, and PowerPoint, producing editable deliverables. Its proficiency in tool utilization is highlighted by:

• Toolathon: 46.3% accuracy, achieving a global top-tier ranking.
• MM Claw: Maintains 97% skill compliance across over 40 complex skills.
• MM Claw end-to-end benchmark: 62.7% accuracy, closely approaching Sonnet 4.6.

Entertainment Features:
The model incorporates enhanced character consistency and emotional intelligence, leading to more engaging interactions in entertainment applications. MiniMax has open-sourced "OpenRoom," an interactive demonstration that integrates AI interaction within a web graphical user interface (GUI) space, offering real-time visual feedback and scene interactions.

Technical Details and Usage:
MiniMax-M2.7 is a substantial model with 229 billion parameters. It supports various tensor types including F32, BF16, and F8\_E4M3, indicating flexibility in precision for deployment. Recommended inference frameworks for serving the model include SGLang, vLLM, Transformers, ModelScope, and NVIDIA NIM. For optimal performance, the recommended inference parameters are $temperature=1.0$, $top_p=0.95$, and $top_k=40$. The default system prompt is "You are a helpful assistant. Your name is MiniMax-M2.7 and is built by MiniMax." Tool calls within the model's prompt are structured using XML format, with $<invoke>$ tags for tool names and $<parameter>$ tags for arguments, such as: