Qwen3.8-27B-Fable-Distill — GGUF

As always, big thank you to @nightmedia for the benchmarks

GGUF conversions of TeichAI/Qwen3.8-27B-Fable-Distill, a BF16 finetune of Qwen3.8-27B (base: Qwen/Qwen3.8-27B) trained with Unsloth + TRL.

The model was trained on a public set of chat and agent traces from Fable 5 as well as a much larger corpus of private personal Fable 5 data.

Converted with llama.cpp b6b4344e.

MTP head kept at BF16

This model ships a multi-token-prediction (nextn) head, and every quant here keeps that head unquantized at BF16 while the other 64 layers are quantized normally:

qwen35.block_count          = 65      # 64 transformer layers + 1 MTP layer
qwen35.nextn_predict_layers = 1
blk.64.*                    = bf16    # 424.7M params, left alone

Files

Usage

Text:

llama-cli -m Qwen3.8-27B-Fable-Distill-Q4_K_M.gguf -c 8192 -p "Hello"

Vision — pass the projector alongside the model:

llama-mtmd-cli -m Qwen3.8-27B-Fable-Distill-Q4_K_M.gguf \
               --mmproj mmproj-F16.gguf \
               --image photo.jpg -p "Describe this image."

Server:

llama-server -m Qwen3.8-27B-Fable-Distill-Q4_K_M.gguf --mmproj mmproj-F16.gguf

Server + MTP:

llama-server -m Qwen3.8-27B-Fable-Distill-Q4_K_M.gguf --mmproj mmproj-F16.gguf --spec-type draft-mtp --spec-draft-n-max 3

Notes

• The model is multimodal (image-text-to-text). Without an mmproj-*.gguf you get a text-only model. Three precisions are provided; F16 is the usual choice, BF16 matches the source weights' dtype, and F32 is there if you want the projector left entirely unquantized.
• Qwen3.5-family chat template with thinking support: it accepts enable_thinking and a reasoning_effort of low, medium or xhigh (the template's own default is xhigh, which thinks at length every turn).
• Base model sampling recommendations: temperature 1.0, top_p 0.95, top_k 20.

The data for this model was easily formatted, validated, and masked using Teich

This qwen3_5 model was trained 2x faster with Unsloth and Huggingface's TRL library.