Qwen3.8 27B DSpark GGUF
GGUF conversion of RadixArk/Qwen3.8-27B-DSpark, a DSpark speculative-decoding drafter for Qwen/Qwen3.8-27B. DSpark support got added very recently in llama.cpp via PR #25173 (merged 2026-07-28).
WARNING: This is highly experimental. I have been unable to get a speedup using DSpark so far. It is very memory hungry for some reason as well. Feel free to experiment with this and see if there are any performance benefits.
Files
| File | Size |
|---|---|
Qwen3.8-27B-DSpark-BF16.gguf | 2.73 GB |
Qwen3.8-27B-DSpark-Q8_0.gguf | 1.46 GB |
Usage
Requires a llama.cpp build containing PR #25173.
# my config for dspark
--spec-draft-hf erlidev/Qwen3.8-27B-DSpark-GGUF:BF16 \
--spec-type draft-dspark \
--spec-draft-n-max 7 \
-ngld 99
Should work with any Qwen3.8 27B quant.
Conversion
Converted with llama.cpp at 9d57ce456c94d241dde672b2db9cf18879766568.
python convert_hf_to_gguf.py ./Qwen3.8-27B-DSpark \
--target-model-dir ./Qwen3.8-27B-tokenizer \
--outtype bf16 --outfile ./Qwen3.8-27B-DSpark-BF16.gguf
llama-quantize Qwen3.8-27B-DSpark-BF16.gguf Qwen3.8-27B-DSpark-Q8_0.gguf Q8_0
Upstream acceptance length
From the source model card:
FP8 target, BF16 draft, block size 7, temp 0.6 / top-k 20 / top-p 0.95.
| Workload | Acceptance length |
|---|---|
| GSM8K | 4.57 |
| MATH-500 | 4.08 |
| MBPP | 3.67 |
| HumanEval | 3.47 |
| AIME 2025 | 3.28 |
| MT-Bench | 3.10 |
| AIME 2026 | 3.07 |
| LBPP | 3.03 |
| Alpaca | 2.95 |
| LiveCodeBench | 2.94 |
| Arena-Hard-v2 | 2.71 |