---

base_model:

• deepseek-ai/dspark_gemma4_12b_block7
• google/gemma-4-12B-it

library_name: gguf

tags:

• llama-cpp
• speculative-decoding
• dspark
• gemma4

license: mit

---

DSpark Gemma4 12B draft (Q4_0 GGUF)

4-bit (pure Q4_0) GGUF draft for DSpark speculative decoding with

google/gemma-4-12B-it.

Converted with llama.cpp ft-dspark from

deepseek-ai/dspark_gemma4_12b_block7.

Files

| File | Description |

|------|-------------|

| dspark_gemma4_12b_q4pure.gguf | Draft model (~1.9 GB), LLM_ARCH_DSPARK |

Requirements

• llama.cpp built from ft-dspark (or a release that includes draft-dspark)
• A separate target Gemma4 12B GGUF (not included here)

Usage

llama-cli \
  -m /path/to/gemma-4-12B-it-Q4_0.gguf \
  -md ./dspark_gemma4_12b_q4pure.gguf \
  --spec-type draft-dspark \
  --spec-draft-n-max 4 \
  -c 512 -ngl 99 -ngld 99 \
  -p "Your prompt" -n 128 --temp 0