Descriptions:
Fahd Mirza tests MiniMax M2.7, a newly released model notable for having participated in its own training process — a self-evolution architecture where the model reads documentation, chains skills, builds memory, runs experiments, and loops results back autonomously under human configuration and steering. MiniMax claims this approach enabled stable performance across complex environments with 50+ tools and 60–150 feature lists, where most models degrade.
The benchmark highlight is M2.7’s performance in 22 Kaggle-style machine learning competitions run over 24 hours autonomously, climbing from a 50% medal rate to nearly 74% and earning 9 gold medals. The video references a head-to-head comparison chart placing M2.7 against Claude Sonnet 4.6, Claude Opus 4.6, and GPT-5.4 across eight real-world task categories spanning coding, tool use, and agentic tasks — with M2.7 leading on several dimensions.
Mirza’s live tests cover: a one-shot interactive HTML/CSS/JavaScript animation (a genie lamp scene with physics, humor, and multilayered interactivity), a comprehensive multilingual translation task across dozens of language families including constructed languages like Klingon and Valyrian with cultural annotations, and a multimodal architectural blueprint analysis estimating suitability for a family of four. Across all three, the model performs well with minimal prompting, and Mirza highlights the unsolicited cultural nuance in the multilingual output as a standout capability.
📺 Source: Fahd Mirza · Published March 18, 2026
🏷️ Format: Review
