Qwen3.5-4B-DFlash-Q4_0-GGUF

Qwen3.5-4B-DFlash Q4_0 (GGUF)

Q4_0 GGUF conversion of z-lab/Qwen3.5-4B-DFlash (DFlash v1 draft model), paired with a Q4_0 GGUF of the Qwen/Qwen3.5-4B base/target model.

Converted with llama.cpp's convert_hf_to_gguf.py (DFlashDraftModel → MODEL_ARCH.DFLASH), then quantized with llama-quantize.

Files

Naming follows the same <prefix>-<base model>.gguf convention llama.cpp MTP sidecars use (e.g. mtp-<model>.gguf), so tooling that expects a draft file next to its base can find it by pattern.

Usage (llama.cpp)

llama-server -m Qwen3.5-4B-Q4_0.gguf -md dflash-Qwen3.5-4B-Q4_0.gguf --spec-type draft-dflash

Notes

• This is DFlash v1: dflash.block_size is set, dflash.selector_top_k is absent. DFlash2 (which adds a selector_top_k candidate-path selector) needs a different checkpoint, e.g. z-lab/Qwen3.8-27B-DFlash2.
• Tested on Hexagon HTP (Snapdragon): the model loads and places all draft layers on HTP, but decode did not complete in testing (stalls after the first op-batch, both with the draft on CPU and on HTP) — see ggml/llama.cpp#369 for details.