---

base_model: z-lab/Qwen3.5-4B-DFlash

tags:

• gguf
• dflash
• speculative-decoding
• draft-model
• qwen3.5

library_name: gguf

---

Qwen3.5-4B-DFlash — Q4_0 GGUF (draft model)

A Q4_0 GGUF of the z-lab/Qwen3.5-4B-DFlash

block-diffusion draft model for DFlash speculative

decoding with the Qwen3.5-4B target. This is not a standalone language model — it must be

paired with the Qwen3.5-4B target in a DFlash-capable llama.cpp server.

Files

| File | Size (bytes) | SHA256 |

| --- | ---: | --- |

| qwen35-4b-dflash-Q4_0.gguf | 367,939,840 | 2772878B6B2B4E607C42BFCEE5A429B54BE27F7A57E86A3108E9E437CB5C79FF |

69 tensors total: 43 2-D weight tensors are Q4_0, the remaining 26 (norms, etc.) are F32.

Provenance

• Source: z-lab/Qwen3.5-4B-DFlash, HF revision 9a1996ccf887b79ab3af4fcbf8c1d1f4b5658bcf

(model.safetensors SHA256 1EB221D36ABB13A5F1B972F8D031A9723FAD8CBB7D275ABE548B60E77577EB42).

• Conversion: upstream llama.cpp convert_hf_to_gguf.py with the DFlash support merged in

ggml-org/llama.cpp#22105, sourcing the

tokenizer from the Qwen3.5-4B target via --target-model-dir. safetensors → BF16 GGUF →

llama-quantize Q4_0.

• The BF16 intermediate is 1,279,873,280 bytes.

Architecture / GGUF metadata

• general.architecture = dflash
• 6 layers: 5 sliding_attention + 1 full_attention (sliding_window = 4096)
• hidden 2560, FFN 9216, 32 Q heads / 8 KV heads, head dim 128
• dflash.block_size = 16
• dflash.target_layers = [2, 6, 10, 14, 18, 22, 26, 30] (llama.cpp layer-input semantics;

the source dflash_config.target_layer_ids are [1, 5, 9, 13, 17, 21, 25, 29])

• tokenizer.ggml.mask_token_id = 248077, vocab size 248320 (shares the Qwen3.5-4B vocab)

The draft is only the DFlash core; at runtime it borrows the target's token embedding and

tied output projection.

Usage (sketch)

Run against the Qwen3.5-4B target in a DFlash-capable llama.cpp server, e.g.:

llama-server \
  -m Qwen3.5-4B-Q4_0.gguf \
  --spec-draft-model qwen35-4b-dflash-Q4_0.gguf \
  --spec-type draft-dflash --spec-draft-n-max 7

DFlash speculative decoding is not part of stock upstream llama.cpp runtime; use a build

with DFlash runtime support.

License

Inherits the terms of the base model — see

z-lab/Qwen3.5-4B-DFlash.