QuantaPlanta/DeepSeek-V4-Flash-Vision-Exp-MXFP4-GGUF overview

---

base_model: deepseek-ai/DeepSeek-V4-Flash-Vision-Exp

---

MXFP4 quant of https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp as it should be, including mmproj and dspark.

$PY $LCPP/convert_hf_to_gguf.py "$SRC" \
--outtype bf16 --no-mtp --split-max-size 45G \
--model-name DeepSeek-V4-Flash-Vision-Exp \
--outfile "$OUT/DeepSeek-V4-Flash-Vision-Exp-MXFP4_MOE.gguf"

$PY $LCPP/convert_hf_to_gguf.py "$SRC" \
  --mmproj --outtype bf16 \
  --outfile "$OUT/mmproj-DeepSeek-V4-Flash-Vision-Exp-BF16.gguf"

$PY $LCPP/convert_hf_to_gguf.py "$SRC" \
  --dspark --target-model-dir "$SRC" --outtype bf16 \
  --outfile "$OUT/dspark-DeepSeek-V4-Flash-Vision-Exp-MXFP4_MOE.gguf"

ATTENTIONE

Blud Claude has a caveat on the DSpark. Please read if you want to use it.

What the workaround costs: the draft loses the vision routing bias, so on image-heavy prompts its expert selection can
diverge from the target's and acceptance drops there. Text-only drafting is unaffected, and correctness never is — the
target verifies every drafted token regardless. Given DSpark was already the speculative part of this plan, that's a fair
trade to get the file built.