Qwen3.8-27B-DFlash2-Exp
An experimental DFlash2 drafter for Qwen/Qwen3.8-27B.
This checkpoint uses a 7-query sample_from_anchor=true layout that produces seven draft proposals from seven query positions instead of eight, reducing draft-side query compute.
Results
| Setting | Value |
|---|---|
| Engine | vLLM |
| Hardware | NVIDIA H200, TP1 |
| Concurrency | 1 |
| Thinking sampling | temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0 |
| Thinking mode | reasoning_effort=xhigh, enable_thinking=true |
| Non-thinking sampling | temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0 |
| Non-thinking mode | enable_thinking=false |
| Max new tokens | 4,096 |
| Context | native 262,144 |
| Draft tokens | 7 per verification step |
| Dataset | Samples | Thinking (xhigh) AL |
Non-thinking AL |
|---|---|---|---|
| AA-LCR | 100 | 3.8475 | 4.3578 |
| AIME26 | 30 | 2.9922 | 5.2214 |
| GSM8K | 1,319 | 3.8272 | 5.5458 |
| HumanEval | 164 | 3.3457 | 6.5176 |
| LiveCodeBench | 1,055 | 2.7684 | 4.5785 |
| LongBench-v2 | 402 | 3.4772 | 3.7079 |
| MATH500 | 500 | 3.4836 | 5.7381 |
| MBPP | 257 | 3.2696 | 5.4883 |
| MMSpec | 600 | 3.1216 | 3.3236 |
| MT-Bench | 80 | 2.8316 | 3.5207 |
| SWE-bench Pro | 731 | 2.6905 | 3.8953 |
Accepted length is calculated from the raw server counters:
accepted_length = 1 + accepted_tokens / draft_calls
Architecture
| Setting | Value |
|---|---|
| Draft layers | 5 |
| Auxiliary hidden-state IDs | [6, 20, 34, 48, 62] |
| Hidden size | 5,120 |
| FFN intermediate | 17,408 |
| Attention | 32 Q heads / 8 KV heads |
| Head dimension | 128 |
| Attention window | sliding window 2,048 |
| Draft attention | non-causal inside the draft block |
| RoPE theta | 1e7 |
| Norm / activation | RMSNorm eps 1e-6 / SiLU |
| Block | 7 draft tokens, sample_from_anchor=true |
| Dynamic convolution | kernel 2, group 16 |
| Candidate selector | rank 256, top-k 16 |
| Draft vocabulary | 248,320 |
| Parameters | 1,924,404,480 |
Training data
A decontaminated, redistributable subset of the training corpus is released as DaoCloud/Qwen3.8-27B-Drafter-SFT.
Example: Run with vLLM
This is a drafter checkpoint, not a standalone language model. Pair it with Qwen3.8-27B or a compatible quantized version:
vllm serve Qwen/Qwen3.8-27B \
--tensor-parallel-size 1 \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder \
--reasoning-parser qwen3 \
--mm-encoder-tp-mode data \
--speculative-config '{
"method": "dflash",
"model": "DaoCloud/Qwen3.8-27B-DFlash2-Exp",
"num_speculative_tokens": 7
}'
This checkpoint uses the 7-query sample_from_anchor=true layout and currently requires vLLM PR #54154.
Checkpoint
The BF16 checkpoint contains 81 tensors and 1,924,404,480 parameters. It is released as a 3,848,817,920-byte (3.584 GiB) safetensors file.
| File | SHA256 |
|---|---|
model.safetensors |
9e1aad9323c3ffecc35b1130e222476f722238233263e7ef9f08389bf1ff2a04 |
Future work
- Adaptive verification: explore confidence-head-driven adaptive draft budgets for high-concurrency serving, including varlen GDN support (vLLM #51869).
- MoE drafter: explore sparsely activated drafter architectures for large-scale serving, increasing model capacity without proportionally increasing active compute.
License
Apache-2.0.