Hacker News
713 HN Comments on Qwen3.8-27B: A Broken Jinja Template Was Costing 25 Points of Agent Success Rate
The 1,191-point HN thread on Qwen3.8-27B is dominated not by benchmarks but by deployment breakage. The community found the shipped Jinja chat template was broken; one developer reported agent success rate jumping from 67% to 92.5% after applying a third-party corrected template. Practitioners also flag severe KV-cache inefficiency — 32K of context consuming 2.5GB of VRAM, with one user unable to fit 128K even quantizing V to Q4_0 — and reasoning-mode overthinking, with a competing 26B-A3B model producing comparable code using roughly a tenth of the thinking tokens.
Source
↳ Follow the thread