DeepSeek-V4-Flash-DSpark-draft-Q3_K_S

DeepSeek V4 Flash DSpark Draft — Q3_K_S

Q3_K_S quantized DSpark draft model for DeepSeek V4 Flash, for use with speculative decoding (DSpark).

Origin

• Source: the DSpark drafter shipped in unsloth/DeepSeek-V4-Flash-0731-GGUF (Q8_0)
• Requantized to Q3_K_S with llama-quantize --allow-requantize
• Size: 8.54 GB

Usage

Use as the draft model together with a DeepSeek V4 Flash main model to accelerate generation.

llama.cpp

llama-server \
  -m <main-model>.gguf \
  --model-draft DeepSeek-V4-Flash-DSpark-draft-Q3_K_S.gguf \
  --spec-type draft-dspark \
  --spec-draft-n-max 2 \
  --spec-draft-type-k q8_0 \
  --spec-draft-type-v q8_0

ds4 (antirez/ds4)

./ds4 -m <main-model>.gguf --rocm --dspark --mtp DeepSeek-V4-Flash-DSpark-draft-Q3_K_S.gguf

Notes

• Q3_K_S is much smaller than the Q8_0 source drafter (~8.5 GB vs ~10 GB) at the cost of slightly lower draft quality / acceptance rate. Benchmark before committing on your hardware.
• The draft model does not need to be the same quant as the main model.

License

MIT — same as upstream unsloth/DeepSeek-V4-Flash-0731-GGUF.

Credits

• unsloth — original DSpark drafter
• antirez/ds4 — DSpark engine
• Kevletesteur — Strix Halo verified GGUF setup that motivated this requantization