Qwen3.8-3.6-27B-blend DFlash2

Unofficial JetBrains derivative. This drafter was prepared by JetBrains by adapting the DFlash 2 model from Inco AI for the Qwen Blend target. The target uses Qwen models developed by Alibaba Cloud. This release is not affiliated with, sponsored by, or endorsed by Alibaba Cloud or Alibaba Group.

Blend collection · DFlash 2 blog · DFlash GitHub · Upstream drafter

This repository contains the DFlash 2 draft model adapted for JetBrains/Qwen3.8-3.6-27B-blend. It runs alongside the target model, proposing tokens for the target to verify, and cannot generate independently.

The target is a 50/50 linear interpolation of Qwen3.6-27B and Qwen3.8-27B. Read Making Local AI Smarter and Faster for the Blend release story, coding evaluations, and MTP runtime experiments.

Related Blend models: BF16 target · GGUF target · MLX 4-bit target · MTP MLX 4-bit drafter.

Model details

DFlash 2 is a block-diffusion drafter: it predicts a block of tokens in parallel, retains candidates at each position, and uses a lightweight selector to choose a coherent path. Two-tap dynamic convolutions help maintain draft quality across the block. The target model verifies the proposed tokens.

Quick start

Splash (Apple silicon)

Splash requires Apple M3 or newer and macOS 26.4 or later. Use the matching MLX target and select this drafter explicitly:

brew install incoai/tap/splash

splash serve \
  --model JetBrains/Qwen3.8-3.6-27B-blend-MLX-4bit \
  --draft-model JetBrains/Qwen3.8-3.6-27B-blend-DFlash2

Splash downloads and prepares both models on first launch, then reuses the prepared weights. It serves OpenAI-compatible Chat Completions and Responses APIs at http://127.0.0.1:8000, with a browser chat at the same address. If Splash is already installed, update it with brew update and brew upgrade incoai/tap/splash.

vLLM (NVIDIA GPU)

DFlash 2 is supported in the standard vLLM release:

pip install -U "vllm>=0.30.0"

vllm serve JetBrains/Qwen3.8-3.6-27B-blend \
  --speculative-config '{
    "method": "dflash",
    "model": "JetBrains/Qwen3.8-3.6-27B-blend-DFlash2",
    "num_speculative_tokens": 7
  }'

SGLang (NVIDIA GPU)

DFlash 2 is supported in the standard SGLang release:

pip install -U "sglang[all]>=0.5.20"

python -m sglang.launch_server \
  --model-path JetBrains/Qwen3.8-3.6-27B-blend \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path JetBrains/Qwen3.8-3.6-27B-blend-DFlash2 \
  --speculative-num-draft-tokens 8

SGLang counts the full block of 8 tokens; vLLM counts its 7 draft proposals.

llama.cpp

The DFlash2 drafter ships as sidecars (dflash-*.gguf) inside the GGUF repository of the blend target, so no separate draft download is needed:

llama serve -hf JetBrains/Qwen3.8-3.6-27B-blend-GGUF --spec-type draft-dflash --spec-draft-n-max 7

DFlash2 support and sidecar discovery are part of upstream llama.cpp (PRs #30548, #30663, #30927) and are included in current prebuilt binaries from llama.app.

MLX (Apple silicon)

Use the DFlash MLX runner with the matching 4-bit target:

pip install -U "dflash[local]"

dflash generate mlx \
  --model JetBrains/Qwen3.8-3.6-27B-blend-MLX-4bit \
  --draft JetBrains/Qwen3.8-3.6-27B-blend-DFlash2 \
  --draft-bits 4 --block-size 5 \
  --reasoning low --temperature 1.0 --top-p 0.95 --top-k 20 \
  --max-new-tokens 2048 \
  "Write a Python function that merges two sorted lists."

--draft-bits 4 quantizes the BF16 drafter at load time. The MLX example uses block size 5, following the DFlash runner's recommendation for quantized models.

For API requests to the servers above, use temperature 1.0, top-p 0.95, and top-k 20, with thinking enabled. Use reasoning_effort: "low" for low-effort reasoning.

Training

This drafter was adapted for JetBrains/Qwen3.8-3.6-27B-blend using agentic trajectories generated with Junie on our internal coding tasks spanning multiple programming languages.

License and attribution

Distributed under the Apache License, Version 2.0.

DFlash 2 was developed by Inco AI. The Qwen target models were developed by Alibaba Cloud. Credit for the original DFlash method belongs to the authors cited below.

Copyright © 2026 JetBrains s.r.o. This notice applies only to original material contributed by JetBrains. Upstream material remains subject to its original copyright and attribution notices.

See NOTICE for attribution and CHANGES.md for a description of the modifications and the affected files.

Note on acceptable use

This model is distributed under the Apache License, Version 2.0. This notice does not modify the License. Users are reminded that using this model, or its outputs, to infringe the copyright or other intellectual property rights of third parties may violate applicable law, independent of the terms of this License.

Citation

JetBrains blog post: [title and URL to be added].

If you use this drafter, please cite DFlash 2 and the original DFlash paper:

@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}
}