Qwen-3.8-27B is good enough that I stopped using API

A user praises Qwen-3.8-27B, stating it's capable of complex refactors unsupervised, making them stop using APIs. They use the Q4_K_S model quant and Q8_0 context, noting Swift-Qwen is faster but prone to loops. The model runs on a Raspberry Pi within a Docker sandbox, utilizing a Pi agent with minimal tools like Bash, read, write, and edit. The edit tool is identified as a weak link due to indentation issues. The user also notes the cost of 1M tokens is comparable to nano-gpt.com.

This user's experience highlights Qwen-3.8-27B's capability to replace API usage for complex refactoring, unlike other models like Swift-Qwen which are faster but prone to loops.