A few objective clarifications: 1) This has nothing to do with MiniMax (I never take sponsored posts). 2) 'Personal experience' does not equal actual performance; it's not quantitative data. After more extensive testing, overall coding ability has seen a qualitative improvement compared to m2.7. However, one downside I've noticed is that the 1-shot results are not as comprehensive as Opus 4.6/4.7/gpt5.5, and there are cases where considerations are not fully thorough. But by adding another Agent as a gate and mentor (e.g., Opus 4.7), you can achieve excellent results. M3 writes the code, Opus 4.7 does code review and provides specific revision suggestions, striking a great balance between 'effectiveness + token usage.' To ensure objectivity, I'm working on a more quantitative evaluation: using real Issues from the Multica project as a benchmark. The entire process is based on Multica's Squads feature, and I'll also include other mainstream OSS models as evaluation targets.