Qwen3.6-27B-DFlash2

Unofficial JetBrains derivative. This drafter was prepared by JetBrains by adapting the DFlash 2 model from Inco AI for Qwen/Qwen3.6-27B, a model developed by Alibaba Cloud. This release is not affiliated with, sponsored by, or endorsed by Alibaba Cloud or Alibaba Group.

This repository contains the DFlash 2 draft model for Qwen3.6-27B, for use with thinking disabled. It runs alongside the target model, proposing tokens for the target to verify, and cannot generate independently. It ships with the Qwen3.6 package of Junie Local.

Previously released as vvsotnikov/Qwen3.6-27B-nothink-DFlash2-v0.1; the weights in this repository are identical.

Drafter lineage: incoai/Qwen3.8-27B-DFlash2 → JetBrains/Qwen3.8-3.6-27B-blend-DFlash2 (blend adaptation) → this Qwen3.6-27B adaptation.

Thinking mode

Use this drafter with thinking disabled on the target. It was fine-tuned on Qwen3.6-27B trajectories generated with thinking disabled; acceptance under thinking-enabled generation has not been measured. Junie Local runs Qwen3.6-27B with thinking off.

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.

Property	Value
Matching target	Qwen/Qwen3.6-27B, revision 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9, thinking disabled
Drafter lineage	incoai/Qwen3.8-27B-DFlash2 → JetBrains/Qwen3.8-3.6-27B-blend-DFlash2 (v0.3, step 500) → this Qwen3.6-27B adaptation
Architecture	DFlash2DraftModel, 5 draft layers
Parameters	1,924,404,480
Stored weights	BF16, Safetensors (81 tensors, 3,848,817,920 bytes)
Weights SHA-256	52d968b9a9f7cc077580327074119a3815f74f48296c5f05b1aa1e0713f7b1f8
Speculation block size	8, corresponding to 7 draft proposals per verification step

Training

Fine-tuned for 100 optimizer updates with a fresh optimizer, starting from the blend adaptation's v0.3 weights (training step 500). The planned 500-update budget was not completed; this release contains the step-100 inference export without optimizer state or training data.

Training uses recorded agentic coding trajectories generated by the target with thinking disabled. Source generation used temperature 0.7, top-p 0.8, top-k 20, and presence penalty 1.5. Token IDs and response loss masks were preserved.

Setting	Value
Updates in this release	100
Effective batch	4 calls: DDP 2, microbatch 1, accumulation 2
Training calls processed	400 (of a 2,000-call pool)
Learning rate	1e-4; 5 warmup updates, then constant
Optimizer	AdamW; FP32 master weights and moments
Draft precision	BF16
Gradient clipping	1
Objective	Recorded-token CE + coupled selector CE (alpha 1), gamma 4
Anchors / objective chunks	64 / 8
Attention backend	SDPA
Target embeddings and output head	Frozen
Seed	42

Training used SpecForge with native SGLang feature extraction and Mooncake transport. Pinned source revisions: SpecForge ed64d275bac8a48e126adb2827368603566a3029; Speculators 3419401a380db305ed6493ba4680ad8a3c3e0e45.

Quick start

vLLM (NVIDIA GPU)

DFlash 2 is supported in the standard vLLM release. Serve with thinking disabled on the target:

pip install -U "vllm>=0.30.0"

vllm serve Qwen/Qwen3.6-27B \
  --speculative-config '{
    "method": "dflash",
    "model": "JetBrains/Qwen3.6-27B-DFlash2",
    "num_speculative_tokens": 7
  }' \
  --chat-template-kwargs '{"enable_thinking":false}'

SGLang (NVIDIA GPU)

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

python -m sglang.launch_server \
  --model-path Qwen/Qwen3.6-27B \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path JetBrains/Qwen3.6-27B-DFlash2 \
  --speculative-num-draft-tokens 8

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

MLX (Apple silicon)

Use the DFlash MLX runner with a matching quantized Qwen3.6-27B target. The drafter quantizes at load time:

pip install -U "dflash[local]"

dflash generate mlx "Write a Python function that checks whether a string is a palindrome." \
  --model <Qwen3.6-27B MLX 4-bit target> \
  --draft JetBrains/Qwen3.6-27B-DFlash2 \
  --draft-bits 4 --block-size 8 \
  --temperature 0.6 --top-p 0.95 --top-k 20 \
  --max-new-tokens 256

Validation status

All 81 exported BF16 tensors were checked for exact equality with the saved training checkpoint. This step-100 checkpoint has not undergone serving benchmarks or downstream agent-quality evaluation. Results from the initializer or separate short training runs do not establish this checkpoint's acceptance rate, speed, or quality.

License and attribution

Distributed under the Apache License, Version 2.0. The original Qwen LICENSE, including Copyright 2026 Alibaba Cloud, is retained unchanged.

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

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

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.