Testing MiniMax M2.7 via API on three real ML and coding workflows
16 hours ago
- #AI Evaluation
- #Workflow Testing
- #MiniMax M2.7
- Tested MiniMax M2.7 API on three ML/coding workflows: refactoring a PyTorch project, drafting Obsidian vault notes, and scaffolding a Kaggle competition entry.
- Compared results with Claude Opus 4.7; M2.7 performed well with explicit constraints and concrete output formats but struggled when context was implicit.
- Refactored an old PyTorch project successfully by providing step-by-step prompts and reviewing changes, highlighting M2.7's suitability for supervised, narrow-scope tasks.
- Drafted and audited Obsidian knowledge notes; M2.7 produced solid first drafts with accurate technical content but required verification of citations and adherence to style rules.
- Scaffolded a Kaggle competition entry; M2.7 handled initial setup and feature engineering but missed Kaggle-specific mechanics (e.g., kernel-only rules) unless explicitly prompted.
- Noted cost and speed advantages: M2.7 processed ~91M tokens at lower cost (~$8) and faster speed (~2x) compared to Opus 4.7, making it suitable for rapid iteration.
- Recommended using M2.7 for supervised refactors, first-draft technical content, audits with explicit criteria, and iterative ML code improvements, but not for open-ended strategy or unverified references.