Qwen3-Coder-30B-A3B-DFlash

Paper (Coming Soon) | GitHub | Blog

DFlash is a novel speculative decoding method that utilizes a lightweight block diffusion model for drafting. It enables efficient, high-quality parallel drafting that pushes the limits of inference speed.

This model is the drafter component. It must be used in conjunction with the target model Qwen/Qwen3-Coder-30B-A3B-Instruct.

📊 Training Data & Efficiency

Qwen3-Coder-30B-A3B-DFlash is trained on 289K samples, composed of:

• Code split from nvidia/Nemotron-Post-Training-Dataset-v2
• theblackcat102/evol-codealpaca-v1
• Approximately 2.8K Cline execution traces collected by ourselves

Despite being trained on significantly less data, DFlash already outperforms EAGLE-3 in inference acceleration. In comparison, lmsys/SGLang-EAGLE3-Qwen3-Coder-30B-A3B-Instruct-SpecForge is trained on the open-perfect-blend dataset with 1.4M samples, nearly 5× more data than DFlash.

This result highlights the training efficiency and scalability of DFlash, and suggests that further scaling the training data can unlock even greater acceleration gains.

🚀 Quick Start

SGLang

DFlash is now supported on SGLang. And vLLM integration is currently in progress.

Installation

uv pip install "git+https://github.com/sgl-project/sglang.git@refs/pull/16818/head#subdirectory=python"

Inference

python -m sglang.launch_server \
    --model-path Qwen/Qwen3-Coder-30B-A3B-Instruct \
    --speculative-algorithm DFLASH \
    --speculative-draft-model-path z-lab/Qwen3-Coder-30B-A3B-DFlash \
    --tp-size 1 \
    --dtype bfloat16 \
    --attention-backend fa3 \
    --mem-fraction-static 0.75 \
    --trust-remote-code \

Transformers

Installation

pip install transformers==4.57.3 torch==2.9.0 accelerate

Inference

import torch
from transformers import AutoModel, AutoModelForCausalLM, AutoTokenizer

model = AutoModel.from_pretrained(
    "z-lab/Qwen3-Coder-30B-A3B-DFlash", 
    trust_remote_code=True, 
    dtype="auto", 
    device_map="cuda:0"
).eval()

target = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-Coder-30B-A3B-Instruct", 
    dtype="auto", 
    device_map="cuda:0"
).eval()

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-Coder-30B-A3B-Instruct")
tokenizer.add_special_tokens({"mask_token": "<|MASK|>"})

prompt = "Please provide a Python implementation of the Bubble Sort algorithm."
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generate_ids = model.spec_generate(
    input_ids=model_inputs["input_ids"], 
    max_new_tokens=2048, 
    temperature=0.0, 
    target=target, 
    mask_token_id=tokenizer.mask_token_id, 
    stop_token_ids=[tokenizer.eos_token_id]
)

print(tokenizer.decode(generate_ids[0], skip_special_tokens=True))

Evaluation

DFlash consistently achieves higher speedups than the state-of-the-art speculative decoding method EAGLE-3. All experiments are conducted using SGLang on a single H200 GPU.

For EAGLE-3, we use the following speculative decoding configuration, which gives the best speedup for EAGLE-3:

• --speculative-num-steps 7
• --speculative-eagle-topk 1
• --speculative-num-draft-tokens 8

For DFlash, we use a block size of 16 during speculation.

We compare against the EAGLE-3 checkpoint
lmsys/SGLang-EAGLE3-Qwen3-Coder-30B-A3B-Instruct-SpecForge,
which is trained on 1.4M samples.

In contrast, DFlash is trained on only 289K samples, yet still delivers superior acceleration. This highlights the training efficiency of DFlash. We expect that scaling the training data further will enable DFlash to achieve even larger speedups.

LiveCodeBench

Batch Size	Method	Output Throughput (tokens/s)	Acceptance Length	Speedup vs. AR
1	Autoregressive	226.39	1.00	1.00×
1	EAGLE3	435.63	4.49	1.92×
1	DFlash	561.71	5.89	2.48×
4	Autoregressive	610.80	1.00	1.00×
4	EAGLE3	1205.69	4.49	1.97×
4	DFlash	1538.73	5.91	2.52×
8	Autoregressive	948.98	1.00	1.00×
8	EAGLE3	1936.89	4.50	2.04×
8	DFlash	2591.36	5.88	2.73×
16	Autoregressive	1456.78	1.00	1.00×
16	EAGLE3	3175.19	4.50	2.18×
16	DFlash	4073.23	5.92	2.80×

---

HumanEval

Batch Size	Method	Output Throughput (tokens/s)	Acceptance Length	Speedup vs. AR
1	Autoregressive	226.75	1.00	1.00×
1	EAGLE3	497.21	5.33	2.19×
1	DFlash	658.93	7.29	2.91×
4	Autoregressive	595.48	1.00	1.00×
4	EAGLE3	1299.63	5.32	2.18×
4	DFlash	1695.39	7.30	2.85×
8	Autoregressive	899.91	1.00	1.00×
8	EAGLE3	1980.40	5.34	2.20×
8	DFlash	2835.01	7.37	3.15×
16	Autoregressive	1362.89	1.00	1.00×
16	EAGLE3	3135.44	5.32	2.30×
16	DFlash	4301.11	7.36	3.16×

Acknowledgement

We are grateful to Yotta Labs for their compute support in training this draft model.

Citation

If you find DFlash useful for your research or applications, please cite our project. The full paper is coming soon!

@article{chen2026dflash,
  title   = {DFlash: Block Diffusion for Flash Speculative Decoding},
  author  = {Chen, Jian and Liu, Zhijian},
  journal = {arXiv preprint},
  year    = {2026},
  url     = {[https://github.com/z-lab/dflash](https://github.com/z-lab/dflash)},
  note    = {Paper coming soon}
}