Fixed and verified in production
quant = "q6k" added to the Qwen3.6-27B catalogue entry, and vram_mb corrected from 54000 (the bf16 checkpoint size) to 42000 (measured q6k residency across…
What a single-GPU fallback would actually cost
Worth stating explicitly, because the acceptance criteria above understate it. ISQ and TP are not two independent conveniences — they are a…
Correction: case B overstates the tensor-parallel cost
The qwen3_moe branch above reads as though TP would be off the table and calls building it "a real piece of work". Both overstate…
Implementation now tracked separately in #251, which carries the decision tree for what happens once model_type is known.
This issue stays as the preparation record — the tooling built, the…
Release slipped to 2026-08-14
The 2026-08-12 18:00 date passed without publication. Qwen/Qwen3.8-27B is now expected a day later.
The watcher has been restarted with a 7-day window (was…