DeepSeek-V4-Flash-0731-DSpark-Drafter-GGUF

DeepSeek V4 Flash 0731 compact DSpark drafters

This repository contains the compact DSpark auxiliary for DeepSeek V4 Flash 0731. It is not a standalone language model and does not include target-model weights.

Artifact

The routed experts use Q2_K. Dense projections use Q8_0. The 81-tensor GGUF uses the standardized llama.cpp general.architecture=dflash schema.

The Q2_K artifact is the smallest option. The mixed artifact uses IQ2_XXS for the gate and up routed-expert tensors and preserves the down routed-expert tensors in native MXFP4. Its remaining quantized tensors use Q8_0. The mixed recipe follows the small-drafter design from antirez/ds4 issue 642 and keeps more expert fidelity while remaining suitable for 128 GiB Apple Silicon Macs. The target, drafter, KV cache, runtime buffers, and other loaded models must still fit together.

Unsloth publishes a full-MXFP4 Q8_0 drafter and a BF16 drafter. This repository provides the smaller mixed expert recipe that Unsloth does not publish.

Download

hf download dev7a/DeepSeek-V4-Flash-0731-DSpark-Drafter-GGUF \
  DeepSeek-V4-Flash-0731-DSpark-Drafter-IQ2_XXS-MXFP4-Q8_0-dflash.gguf

Provenance and verification

The source is deepseek-ai/DeepSeek-V4-Flash-0731 revision 9e165c30e2704aec5d9d593cce3eebd58bbef1cb. The repository contains the pinned source manifest, the Q2_K converter, standardized dflash rewriters, verification scripts, tests, build manifests, and checksums.

Both releases passed these schema checks:

• 81 canonical dflash tensors with 32-byte alignment.
• Byte-identical tensor payloads before and after the schema rewrite.
• Byte-identical repeated builds.
• Complete target-tokenizer metadata copied through the official llama.cpp DeepSeek V4 DSpark converter at commit 15586e2d7165570fb3aa7c26e0d442e289ef69de, including tokenizer.ggml.mask_token_id=128799.
• A real llama.cpp draft-dspark decode generated 120 draft tokens and accepted 38, with no invalid -1 token.
• ns4 build e470f38280d9ff4f020c242233998562a88463ee loaded the corrected Q2 file and completed an exact 32-token decode, drafting 47 and accepting 12.

The Q2_K build also passed its numeric limits: maximum Q2_K block-relative error 0.49072265625, below 0.5, and maximum Q8_0 block-relative error 0.00396728515625, below 0.01.

The mixed artifact was converted with antirez/ds4 pull request 645 at commit d147270d7422ba0eb55985c8d4075b9dfaa84bf1. It uses the public routed-MoE importance matrix from antirez/deepseek-v4-gguf. The build manifest pins its revision, checksum, and the deterministic MTP alias transformation.

Reproduce the artifact on Linux AArch64 with Python 3.14:

uv venv --python 3.14.6 .venv
uv pip install --python .venv/bin/python \
  --require-hashes --only-binary=:all: \
  -r requirements-linux-aarch64-py314.lock
.venv/bin/python scripts/download_sources.py --destination sources
.venv/bin/python -m unittest discover -s tests -v
.venv/bin/python scripts/reproduce.py \
  --sources sources --recipe q2_k-q8_0 \
  --manifest-dir manifest --repeat-check
.venv/bin/python scripts/dflash.py \
  --sources sources --recipe q2_k-q8_0 \
  --legacy-input DeepSeek-V4-Flash-0731-DSpark-Drafter-Q2_K-Q8_0.gguf \
  --target-tokenizer-gguf tokenizer.gguf \
  --manifest-dir manifest --repeat-check

# The mixed legacy GGUF is produced with the pinned ds4 converter and imatrix.
.venv/bin/python scripts/rewrite_mixed_dflash.py \
  DeepSeek-V4-Flash-0731-DSpark-Drafter-IQ2_XXS-MXFP4-Q8_0.gguf \
  --target-tokenizer-gguf tokenizer.gguf \
  --manifest-dir manifest

Compatibility

Use this file only with a compatible DeepSeek V4 Flash 0731 target. The runtime must support the standardized llama.cpp dflash GGUF schema and the selected Q2_K or IQ2_XXS/MXFP4 expert recipe. This is a community conversion, not an official DeepSeek release.

The source-derived weights remain under DeepSeek's MIT license. Conversion code and third-party notices are provided in LICENSE.code and THIRD_PARTY_NOTICES.md.