I came across an interesting discussion in several AI-related communities about this Reddit post: one user decided to test how the local Qwen 3.8 27B model behaves in two different agent shells — PI Agent and OpenCode.
From his observations, PI Agent delivers higher-quality results, uses fewer tokens, runs faster and more stably, and most importantly — it doesn’t have a hard 32k output limit.
What impressed me most was context compression: in OpenCode it starts kicking in at around 67k (with a total context window of 100k), while in PI Agent the model “holds up” until 90k, even if you set output to 64k and above.
For those who want to reproduce this locally — even an RTX 3090 is enough, plus a small fine-tune of the model itself (you can find that in the original post).
At the same time, one of the commenters already staged a “duel” between local Qwen paired with PI Agent and Claude Code. The results were very close, and after a small refinement, the PI version even slightly outperformed Claude.
