AnalysisAI ModelsSeptember 12, 2026

User compares Qwen-Next and 3.8 27B for coding on M5 Max

Reddit user runs both models locally on an M5 Max with 128GB RAM, using the q8 27B build and MTPLX's speed-optimized dynamic q4 with 8-bit attention. They report Qwen-Next feels worse than 3.8 27B for coding in a Node.js pi agent, but suspect they're missing something.

1 source

More stories today

Open the live feed
User compares Qwen-Next and 3.8 27B for coding on M5 Max