Instructions to use wyattearp/Qwen3.8-27B-DFlash2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use wyattearp/Qwen3.8-27B-DFlash2 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="wyattearp/Qwen3.8-27B-DFlash2")

# Load model directly
from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("wyattearp/Qwen3.8-27B-DFlash2")
model = AutoModel.from_pretrained("wyattearp/Qwen3.8-27B-DFlash2", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use wyattearp/Qwen3.8-27B-DFlash2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "wyattearp/Qwen3.8-27B-DFlash2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "wyattearp/Qwen3.8-27B-DFlash2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/wyattearp/Qwen3.8-27B-DFlash2

SGLang

How to use wyattearp/Qwen3.8-27B-DFlash2 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "wyattearp/Qwen3.8-27B-DFlash2" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "wyattearp/Qwen3.8-27B-DFlash2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "wyattearp/Qwen3.8-27B-DFlash2" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "wyattearp/Qwen3.8-27B-DFlash2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use wyattearp/Qwen3.8-27B-DFlash2 with Docker Model Runner:
```
docker model run hf.co/wyattearp/Qwen3.8-27B-DFlash2
```

Qwen3.8-27B-DFlash2

Research Blog | GitHub Repository

This repository provides the DFlash 2 Speculative Draft Model for Qwen/Qwen3.8-27B and its derivatives, including JonathanColetti/Qwen3.8-27B-Uncensored and wyattearp/Qwen3.8-27B-Uncensored-NVFP4.

Important: This is a speculative decoding draft model, not a standalone generative LLM. It is designed to run concurrently inside an inference server (vLLM or SGLang) alongside the target base model, proposing draft token blocks that the target model verifies in parallel.

How DFlash 2 Works

DFlash 2 is a block-diffusion speculative drafter:

Parallel Block Proposal: Instead of autoregressively drafting tokens one-by-one, DFlash 2 predicts a full block of 7 tokens in a single forward pass.
Candidate Path Selector: It retains the top candidates at each position and uses a lightweight selector to trace a coherent path through them.
Lossless Verification: The target model (e.g. Qwen3.8-27B-Uncensored-NVFP4) computes verification logits across the proposed block in a single step. Greedy decoding matches the base model with 100% mathematical fidelity, while sampling preserves the target distribution.

Compatible Target Base Models

Target Model	Format / Precision	Hugging Face Repository
Qwen 3.8-27B-Uncensored (NVFP4)	NVFP4 (26.6 GB)	`wyattearp/Qwen3.8-27B-Uncensored-NVFP4`
Qwen 3.8-27B-Uncensored (BF16)	BF16 (54.0 GB)	`JonathanColetti/Qwen3.8-27B-Uncensored`
Qwen 3.8-27B (Standard NVFP4)	NVFP4 (26.6 GB)	`wyattearp/Qwen3.8-27B-NVFP4`
Qwen 3.8-27B (Standard BF16)	BF16 (54.0 GB)	`Qwen/Qwen3.8-27B`

Serving Instructions

1. High-Throughput Serving with vLLM (NVFP4 Base + DFlash 2)

# Install vLLM with DFlash speculative decoding support
pip install -U "vllm>=0.27.1"

# Serve the NVFP4 Uncensored base model paired with this DFlash 2 draft head
vllm serve wyattearp/Qwen3.8-27B-Uncensored-NVFP4 \
  --host 0.0.0.0 \
  --port 8000 \
  --dtype auto \
  --trust-remote-code \
  --gpu-memory-utilization 0.75 \
  --max-model-len 16384 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml \
  --reasoning-parser qwen3 \
  --speculative-config '{
    "method": "dflash",
    "model": "wyattearp/Qwen3.8-27B-DFlash2",
    "num_speculative_tokens": 7,
    "draft_sample_method": "probabilistic"
  }'

2. Serving with SGLang

pip install "sglang[all] @ git+https://github.com/sgl-project/sglang.git#subdirectory=python"

python -m sglang.launch_server \
  --model-path JonathanColetti/Qwen3.8-27B-Uncensored \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path wyattearp/Qwen3.8-27B-DFlash2 \
  --speculative-num-draft-tokens 8

Empirical Verification on DGX Spark (NVIDIA GB10)

Benchmarked when paired with wyattearp/Qwen3.8-27B-Uncensored-NVFP4 across 327 verified evaluation samples:

Decompilation Pass@1 Accuracy: 95.0% (19/20 passed programmatic clang++ -O2 compilation + unit test assertion execution).
Magma ASan Bug Triage: 100.0% (15/15 CVE root causes accurately localized across OpenSSL, SQLite, PHP, PNG, TIFF).
Speculative Verification: Proposes 7 tokens per verification step with zero accuracy regression compared to non-speculative autoregressive decoding.

Citations

If you utilize this DFlash 2 draft checkpoint, please cite the underlying research, fine-tune, and base models:

Inco AI & DFlash 2

@misc{inco2026dflash2,
  title  = {{DFlash 2: Keep Drafting Parallel}},
  author = {{Inco AI}},
  year   = {2026},
  month  = {August},
  url    = {https://inco.ai/blog/dflash2/}
}

@inproceedings{chen2026dflash,
  title     = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
  author    = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}

Jonathan Coletti (Uncensored Fine-Tune)

@misc{coletti2026qwen38uncensored,
  title  = {{Qwen3.8-27B-Uncensored: Uncensored Fine-Tune for Security & Technical Research}},
  author = {Coletti, Jonathan},
  year   = {2026},
  url    = {https://huggingface.co/JonathanColetti/Qwen3.8-27B-Uncensored}
}

Qwen Team

@article{qwen2025qwen25,
  title   = {{Qwen2.5 Technical Report}},
  author  = {{Qwen Team}},
  journal = {arXiv preprint arXiv:2412.15115},
  year    = {2024}
}