llama.cpp quantizations of z-lab/Qwen3.6-35B-A3B-DFlash, the block-diffusion drafter for DFlash speculative decoding. Pair it with Qwen/Qwen3.6-35B (or a quant of it).

Simalar to spiritbuun/Qwen3.6-27B-DFlash-GGUF, but for Qwen3.6-35B-A3B.

Requirements

DFlash speculative decoding is not yet in upstream llama.cpp. You need the fork:

Fork: spiritbuun/buun-llama-cpp (branch master)
SWA support for the DFlash drafter landed in commit b9d01582b (SD-073). Older checkpoints will load the drafter but produce garbage.
Built with: cmake -B build -DGGML_CUDA=ON -DGGML_NATIVE=ON -DGGML_CUDA_FA=ON -DGGML_CUDA_FA_ALL_QUANTS=ON

Usage

`llama-server`

bash

./build/bin/llama-server \
    -m   /path/to/Qwen3.6-35B-target.Q4_K_M.gguf \
    -md  /path/to/dflash-draft-3.6-q8_0.gguf \
    --spec-type dflash \
    -ngl 99 -ngld 99 \
    -np 1 -c 6048 -cd 256 \
    -fa on -b 256 -ub 64 \
    --host 0.0.0.0 --port 8080 --jinja \
    --chat-template-kwargs '{"enable_thinking": false}'

Thinking footgun: the Qwen3.6 chat template enables <think>…</think> by default. That collapses DFlash acceptance because the drafter wasn't trained on the think-wrapped distribution. Pass --chat-template-kwargs '{"enable_thinking": false}' to disable it (≈1.8× throughput uplift).

`llama-speculative-simple`

bash

./build/bin/llama-speculative-simple \
    -m   /path/to/Qwen3.6-35B-target.Q4_K_M.gguf \
    -md  /path/to/dflash-draft-3.6-q8_0.gguf \
    --spec-type dflash \
    -ngl 99 -ngld 99 \
    -c 4096 --draft-max 16 --draft-min 1 \
    -p "Write a Python mergesort."

Note on comparison with the 3.5 drafter

Short-context code prompts do not exercise the sliding-window attention (most queries fall inside the 2048-token window anyway), so the 3.6 drafter's architectural change doesn't produce a dramatic win on this benchmark. The SWA infrastructure is expected to matter on longer-context workloads (> 2 k generated tokens). On short code, Q8_0 on 3.6 is ≈1.3× the throughput of Q4_K_M on 3.5 because the 3.6 target pairs slightly better with the retrained drafter.

Quantization details

Source: z-lab/Qwen3.6-35B-A3B-DFlash (BF16 safetensors, 0.5 B parameters)
Converter: convert_hf_to_gguf.py from spiritbuun/buun-llama-cpp — emits qwen35.attention.sliding_window + qwen35.attention.sliding_window_pattern so the runtime builds per-layer SWA masks

Reproducing the conversion

Tokenizer heads-up: the upstream z-lab/Qwen3.6-35B-A3B-DFlash repo ships only config.json, model.safetensors, and a README — no tokenizer files. The drafter shares the target model's tokenizer. Copy the Qwen3.6 tokenizer files into the drafter directory first.

bash

# 1. Pull the DFlash drafter weights
hf download z-lab/Qwen3.6-35B-A3B-DFlash --local-dir ./dflash-drafter-3.6

# 2. Pull tokenizer files from the target model into the same directory
hf download Qwen/Qwen3.6-35B-A3B-DFlash \
    tokenizer.json tokenizer_config.json vocab.json merges.txt \
    special_tokens_map.json \
    --local-dir ./dflash-drafter-3.6

# 3. Convert to GGUF (F16 first, then quantize)
python convert_hf_to_gguf.py ./dflash-drafter-3.6 \
    --outtype f16 \
    --outfile dflash-draft-3.6-f16.gguf

# 4. Quantize
./build/bin/llama-quantize dflash-draft-3.6-f16.gguf dflash-draft-3.6-q8_0.gguf Q8_0
./build/bin/llama-quantize dflash-draft-3.6-f16.gguf dflash-draft-3.6-q4_k_m.gguf Q4_K_M

Required files in ./dflash-drafter-3.6/ before step 3:

File	Source
`config.json`	`z-lab/Qwen3.6-35B-A3B-DFlash` (has `architectures: ["DFlashDraftModel"]`, `use_sliding_window: true`, `layer_types: [...]`)
`model.safetensors`	`z-lab/Qwen3.6-35B-A3B-DFlash`
`tokenizer.json`, `tokenizer_config.json`, `vocab.json`, `merges.txt`	`Qwen/Qwen3.6-35B`

The converter auto-detects DFlashDraftModel from config.json and emits the SWA metadata when use_sliding_window is set.

Original model card — `z-lab/Qwen3.6-35B-A3B-DFlash`

license: mit library_name: transformers pipeline_tag: text-generation tags:

dflash
speculative-decoding
block-diffusion
draft-model
efficiency
qwen
diffusion-language-model

Qwen3.6-35B-A3B-DFlash

Paper | GitHub | Blog

DFlash is a speculative decoding method that uses a lightweight block diffusion model to draft multiple tokens in parallel. This is the drafter model, which must be paired with Qwen/Qwen3.6-35B-A3B.

Quick Start

Installation

vLLM (We temporarily modify the installation through this PR to support interleaved SWA and ensure correct handling of target hidden states for optimal performance):

bash

uv pip install vllm
uv pip install -U --torch-backend=auto "vllm @ git+https://github.com/vllm-project/vllm.git@refs/pull/40898/head"

SGLang:

uv pip install "git+https://github.com/sgl-project/sglang.git@refs/pull/20547/head#subdirectory=python"

Launch Server

vLLM:

bash

vllm serve Qwen/Qwen3.6-35B-A3B \
  --speculative-config '{"method": "dflash", "model": "z-lab/Qwen3.6-35B-A3B-DFlash", "num_speculative_tokens": 15}' \
  --attention-backend flash_attn \
  --max-num-batched-tokens 32768

SGLang:

bash

# Optional: enable schedule overlapping (experimental, may not be stable)
# export SGLANG_ENABLE_SPEC_V2=1
# export SGLANG_ENABLE_DFLASH_SPEC_V2=1
# export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1

python -m sglang.launch_server \
    --model-path Qwen/Qwen3.6-35B-A3B \
    --speculative-algorithm DFLASH \
    --speculative-draft-model-path z-lab/Qwen3.6-35B-A3B-DFlash \
    --speculative-num-draft-tokens 16 \
    --tp-size 1 \
    --attention-backend fa3 \
    --mem-fraction-static 0.75 \
    --mamba-scheduler-strategy extra_buffer \
    --trust-remote-code

Tip: For long-context or agentic workloads, add --speculative-dflash-draft-window-size WINDOW_SIZE to enable sliding-window attention for the drafter.

Usage

python

from openai import OpenAI

client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="Qwen/Qwen3.6-35B-A3B",
    messages=[{"role": "user", "content": "Write a quicksort in Python."}],
    max_tokens=4096,
    temperature=0.0
)
print(response.choices[0].message.content)

Benchmark Results

Setup: Single NVIDIA B200, SGLang, thinking enabled, max output length 4096. We report end-to-end throughput, including prefill time. See our GitHub repository for reproduction scripts.

Throughput and Speedup

DFlash achieves up to 2.9x speedup at concurrency 1.

Tokens/sec (speedup vs. autoregressive baseline)

Block Size = 16

Task	Concurrency	AR	DFlash
Math500	1	234	682 (2.9x)
	8	1266	3138 (2.5x)
	16	1954	4813 (2.5x)
	32	2755	6520 (2.4x)
GSM8K	1	235	556 (2.4x)
	8	1236	2564 (2.1x)
	16	1886	3821 (2.0x)
	32	2699	5239 (1.9x)
HumanEval	1	238	603 (2.5x)
	8	1255	2800 (2.2x)
	16	1944	4208 (2.2x)
	32	2767	5782 (2.1x)
MBPP	1	235	559 (2.4x)
	8	1224	2538 (2.1x)
	16	1948	3816 (2.0x)
	32	2780	5378 (1.9x)
MT-Bench	1	233	442 (1.9x)
	8	1238	2028 (1.6x)
	16	1885	2997 (1.6x)
	32	2633	4034 (1.5x)
Alpaca	1	235	393 (1.7x)
	8	1221	1782 (1.5x)
	16	1844	2567 (1.4x)
	32	2579	3689 (1.4x)

Block Size = 8

Task	Concurrency	AR	DFlash
Math500	1	234	617 (2.6x)
	8	1266	2839 (2.2x)
	16	1954	4465 (2.3x)
	32	2755	6614 (2.4x)
GSM8K	1	235	540 (2.3x)
	8	1236	2466 (2.0x)
	16	1886	3899 (2.1x)
	32	2699	5713 (2.1x)
HumanEval	1	238	561 (2.4x)
	8	1255	2655 (2.1x)
	16	1944	4135 (2.1x)
	32	2767	6059 (2.2x)
MBPP	1	235	497 (2.1x)
	8	1224	2324 (1.9x)
	16	1948	3636 (1.9x)
	32	2780	4884 (1.8x)
MT-Bench	1	233	438 (1.9x)
	8	1238	2060 (1.7x)
	16	1885	3182 (1.7x)
	32	2633	4720 (1.8x)
Alpaca	1	235	407 (1.7x)
	8	1221	1880 (1.5x)
	16	1844	2903 (1.6x)
	32	2579	4115 (1.6x)

Acceptance Length

Task	B8	B16
Math500	5.56	7.35
GSM8K	5.21	6.73
HumanEval	5.09	6.44
MBPP	4.78	5.83
MT-Bench	4.20	5.14
Alpaca	3.94	4.62

Acknowledgements

Special thanks to David Wang for his outstanding engineering support on this project. We are also grateful to Modal, InnoMatrix, and Yotta Labs for providing the compute resources used to train this draft model.

Citation

If you find DFlash useful, please cite our work. To share feedback on DFlash or request new model support, please fill out this form: DFlash Feedback.

bibtex

@article{chen2026dflash,
  title   = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
  author  = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  journal = {arXiv preprint arXiv:2602.06036},
  year    = {2026}
}

starskyzheng/Qwen3.6-35B-DFlash-GGUF

Qwen3.6-35B-DFlash — GGUF (Q8_0)

Requirements

Usage

llama-server

llama-speculative-simple

Note on comparison with the 3.5 drafter

Quantization details

Reproducing the conversion

Original model card — z-lab/Qwen3.6-35B-A3B-DFlash

Qwen3.6-35B-A3B-DFlash

Quick Start

Installation

Launch Server

Usage

Benchmark Results

Throughput and Speedup

Acceptance Length

Acknowledgements

Citation

starskyzheng/
Qwen3.6-35B-DFlash-GGUF

`llama-server`

`llama-speculative-simple`

Original model card — `z-lab/Qwen3.6-35B-A3B-DFlash`