Qwen3.6-27B Fable Fusion 711
by DavidAU · Model · Language & code models
Multi-stage open fine-tune of Qwen3.6 27B with refusals ablated, shipped as GGUF quants with built-in MTP heads for consumer GPUs
Evidence
- Aug 7Multilingual translation test got stuck in a loop and hallucinated (e.g. Hausa); reviewer calls multilinguality broken in the quant
- Aug 7Claimed to beat base Qwen3.6 27B on six of seven benchmarks and match it on one; beats a 35B model on all seven
- Aug 78-bit on an A6000 under llama.cpp: 43.6 tok/s with MTP vs 22 without; draft acceptance 62%. Model card claimed about 20% gain