ByteShape released ShapeLearn-quantized GGUF versions of Qwen3.8-27B. Five GPU-optimized variants range from 2.56 to 3.84 bits per weight. Each passed a multi-benchmark evaluation against the BF16 baseline. The models include an embedded MTP head for speculative decoding. DFlash 2 offers the fastest option but requires a separate draft model. Vision capability is supported with MTP but not with DFlash 2. Throughput gains from speculative decoding vary by model and workload. Recommended sampling parameters differ for thinking and instruct modes. The release replaces prior ShapeLearn-Lite quantizations.
What changed
Earlier ShapeLearn-Lite quantizations of Qwen3.8-27B were used.
What it unlocks
Running Qwen3.8-27B with lower memory and higher speed on consumer GPUs.
- 2.56 to 3.84 bits per weight
- 8.8 to 13.1 GB model size
- 1.34–2.10x throughput with DFlash 2
- 832,621 downloads last month
What you need to act on it
- llama.cpp build b10658 or newer for DFlash 2
- GPU with sufficient VRAM for chosen model
Sources