Shapelearn Qwen 3.8 27B (13.1 GB VRAM)
Summary
ByteShape's ShapeLearn blog announces the full ShapeLearn release for Qwen 3.8 27B, including the ShapeLearn-Lite and full ShapeLearn models and their GGUFs. The article presents extensive multi-GPU benchmarks (6 GPUs) showing full ShapeLearn moving the frontier, with GPU-5 as the default choice when memory permits, and discusses decoding methods (MTP and DFlash2) and memory tradeoffs. It also notes that KLD metrics do not directly map to task performance and outlines methodology.