Qwen3.8-27B-DFlash2

Qwen3.8-27B-DFlash2

A draft model that speeds up Qwen3.8-27B by proposing a block of tokens for the main model to check. Not standalone. For people serving Qwen3.8-27B with SGLang or vLLM.

Made by z-lab on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Qwen3.8-27B-DFlash2 is a small helper model, called a drafter, that pairs with Qwen/Qwen3.8-27B. It is not a language model by itself. A speculative decoding server loads it next to the main model. It guesses upcoming tokens, and the main model checks them. The repository mirrors a copy published by Inco AI.

Who it suits: People who already serve Qwen3.8-27B and want faster generation without changing the answers.

Industry | Who uses it | How they use it | Example
Software | Inference engineers | Cut latency on a served Qwen3.8-27B | A coding assistant that replies sooner
Platform | Serving teams | Raise tokens per second under load | A chat service handling several users
Research | Systems researchers | Compare drafting methods | MTP versus DSpark versus DFlash 2

You can skip it if you do not run Qwen3.8-27B, since the draft only works with that target model.

Where to use it:

• SGLang, with the DFLASH speculative algorithm
• vLLM, with a dflash speculative config
• A GPU server, since the card measured on one NVIDIA H200

Use case | Who | What they get back
Faster single-user decoding | Developers | More output tokens per second than plain decoding
Faster batched serving | Platform teams | Higher throughput at moderate concurrency
Lossless drafting | Researchers | Output that matches the target model

How to set it up:

• A GPU server that can serve Qwen/Qwen3.8-27B
• SGLang or vLLM, installed as the card describes
• The draft model, which the server fetches or you download

1. Install SGLang or the vLLM build the card names, using its pip line.
2. Start the server with Qwen3.8-27B as the main model.
3. Add the DFlash 2 draft model through the speculative options.

pip install -U huggingface_hub
hf download z-lab/Qwen3.8-27B-DFlash2

How to use it: Start the server with the main model and the draft model together, then send requests as usual. The card gives one command for SGLang and one for vLLM, and says the output is lossless.

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

vllm serve Qwen/Qwen3.8-27B \
  --speculative-config '{
    "method": "dflash",
    "model": "incoai/Qwen3.8-27B-DFlash2",
    "num_speculative_tokens": 7
  }'

Example: You start the SGLang server with those flags and chat with Qwen3.8-27B as usual. The answers match plain decoding, and the card reports higher output tokens per second.

For newcomers and experts: If you are new, know that this file does nothing alone. It is a helper that sits beside Qwen3.8-27B inside the server, so set up the main model first.

If you know the field, DFlash 2 is a block-diffusion drafter. In one forward pass it proposes several tokens together, with a few top options per position. A selector then traces one path through them. The card says greedy output matches the target exactly and sampling keeps its distribution.

Draft model proposes a block of tokens
Selector picks one coherent path
Qwen3.8-27B verifies the tokens
Accepted tokens become output

Title: Output tokens per second on GSM8K, one request at a time
Plain decoding: 68.9 tok/s
MTP: 178.5 tok/s
DSpark: 185.3 tok/s
DFlash 2: 236.1 tok/s

Good to know:

• License: Apache 2.0
• Cost to run: your own GPU server, as the card measured on an NVIDIA H200
• Data it reaches: the files you download and the prompts you send your own server
• Not for: standalone use, since it only drafts tokens for Qwen3.8-27B
• Limits: at 32 concurrent requests the gains shrink, and the card shows near 1.0 on one task

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 1.9B
• Weights: about 3.8 GB in BF16; about 962 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 46,660
• Likes: 335
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 22
• Fine-tunes: 15
• Tags: transformers, safetensors, qwen3, dflash2, speculative-decoding, block-diffusion, draft-model, sglang, vllm, text-generation

Links

• Model: https://huggingface.co/z-lab/Qwen3.8-27B-DFlash2
• Files: https://huggingface.co/z-lab/Qwen3.8-27B-DFlash2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download z-lab/Qwen3.8-27B-DFlash2

Then follow the usage in its model card. Nothing in this kit runs by itself.