license: mit
library_name: transformers
pipeline_tag: text-generation
tags:

• deepseek
• deepseek-v4
• moe
• mixed-expert
• fp4
• fp8
• quantization
• safetensors
• abliterated
  base_model: huihui-ai/Huihui-DeepSeek-V4-Flash-0731-abliterated-GGUF
  inference:
  parameters:
  do_sample: true

DeepSeek-V4-Flash Q4-MXFP4 — huihui-ai GGUF converted to safetensors + DSpark head

This is a conversion of the DeepSeek-V4-Flash Q4-MXFP4 GGUF from
hf://huihui-ai/Huihui-DeepSeek-V4-Flash-0731-abliterated-GGUF/DeepSeek-V4-Flash-Q4-mxfp4-0731.gguf
to native safetensors format, with the DSpark 3-stage speculative drafter head
included on top.

• Source: huihui-ai/Huihui-DeepSeek-V4-Flash-0731-abliterated-GGUF
  → DeepSeek-V4-Flash-Q4-mxfp4-0731.gguf (Q4-MXFP4 quant, abliterated),
  converted to safetensors shards.
• Main model: 43 layers in model-00001..00045-of-00048.safetensors
  (routed experts as native packed FP4, other quantized matrices as 128x128
  E4M3/E8M0 block FP8).
• DSpark drafter: model-00046..00048-of-00048.safetensors
  (mtp.0 -> 46, mtp.1 -> 47, mtp.2 -> 48), mirroring the official repo's
  placement. Routed experts retain source MXFP4 values losslessly; other
  matrices use block FP8. The drafter weights are an additional inclusion,
  sourced from a separate DeepSeek-V4-Flash-DSpark draft GGUF (see
  conversion_manifest.json); they are not part of the huihui-ai GGUF.
• config.json is the official DSpark config verbatim, laid out to load
  exactly like deepseek-ai/DeepSeek-V4-Flash-DSpark.
• inference/config.json carries the official loader's n_mtp_layers: 3.

This is a format conversion of already-quantized weights, not a recovery of the
original FP8 checkpoint. See conversion_manifest.json for provenance.