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

DeepSeek V4 Flash 0731 DSpark Drafter GGUF

This repository contains only the DSpark auxiliary drafter extracted from DeepSeek V4 Flash 0731. It is not a standalone language model and does not contain the target model weights.

The GGUF is built from the official deepseek-ai/DeepSeek-V4-Flash-0731 checkpoint using the standalone Python scripts included here. The complete source manifest, build recipe, structural parser, numeric verifier, dependency pin, and build results are published alongside the artifact.

Artifact

File	Bytes	SHA-256
DeepSeek-V4-Flash-0731-DSpark-Drafter-MXFP4-Q8_0.gguf	10,897,110,272	5c8a1ce75e4797c170e22ac65080281a1b694c77f2e4fa7a703fbf09c5de0e33
DeepSeek-V4-Flash-0731-DSpark-Drafter-Q2_K-Q8_0.gguf	6,971,241,728	a3f1156e1e4d2f46e1d4367b1bd09e6923adee61aecd41e880e233df33d30a7b
DeepSeek-V4-Flash-0731-DSpark-Drafter-Q2_K-Q8_0-dflash.gguf	6,971,243,008	3e2be643b7881ac61e49c9907a963bdbbfcffe89c4d15c5f0e99e827e0305914

The first two artifacts use GGUF v3 with the custom architecture deepseek_v4_flash_dspark_draft and contain the same 81 tensors. They differ only in the encoding of the nine routed-expert tensors:

Output type	Tensors	Transformation
F32	45	BF16 expanded exactly to F32, or source F32 bytes preserved
F16	2	BF16 Markov heads rounded to F16
Q8_0	25	FP8 E4M3 weights with E8M0 block scales dequantized, then quantized to Q8_0
MXFP4	9	Routed experts in the mxfp4-q8_0 artifact: FP4 nibbles and E8M0 scales repacked without requantization
Q2_K	9	Routed experts in the q2_k-q8_0 artifact: source MXFP4 decoded, then quantized to Q2_K

The MXFP4 transformation is bit-preserving. Q2_K, Q8_0, and F16 transformations are lossy and are checked against explicit numeric error bounds. This repository does not claim that either complete conversion is lossless.

Pinned source

• Repository: deepseek-ai/DeepSeek-V4-Flash-0731
• Revision: 9e165c30e2704aec5d9d593cce3eebd58bbef1cb
• Input: only safetensors shards 46, 47, and 48
• Total source size: 10,863,342,388 bytes

Exact file sizes and SHA-256 values are in manifest/source.json. The downloader and converter refuse any source that does not match them.

Reproduce

The reference build used Linux aarch64, Python 3.14.6, NumPy 2.5.1, uv 0.11.29, and Hugging Face CLI 1.24.0. The NumPy wheel is pinned by hash in requirements-linux-aarch64-py314.lock. The upstream checkpoint is public, so a Hugging Face token is not required for source download.

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
sha256sum -c SHA256SUMS

Use --recipe mxfp4-q8_0 to reproduce the MXFP4/Q8_0 artifact instead.

--repeat-check performs a second clean conversion and requires its complete GGUF SHA-256 to match the verified first build byte for byte. The converter memory-maps the source shards and processes FP8 tensors in bounded bands. It does not require PyTorch, CUDA, safetensors, a GGUF package, or a GPU.

Verification

The included verifier checks pinned source hashes and tensor inventories, GGUF structure and metadata, exact F32/F16 bytes, exact MXFP4 repacking, Q2_K reconstruction error, and Q8_0 encoding plus reconstruction error against a fresh source decode.

Reference-build results:

Check	Result
Tensor count	81
Exact F32/F16 tensors	47
Exact MXFP4 tensors	9
Q2_K tensors	9
Maximum Q2_K block-relative error	0.49072265625
Q2_K limit	0.5
Maximum Q8_0 block-relative error	0.00396728515625
Q8_0 limit	0.01
Repeat build	byte-identical

Machine-readable observations are in the MXFP4/Q8_0 validation and build manifests, and the Q2_K/Q8_0 validation and build manifests.

Standardized dflash migration candidate

DeepSeek-V4-Flash-0731-DSpark-Drafter-Q2_K-Q8_0-dflash.gguf is the llama.cpp-standardized migration candidate for ns4 issue #205. It uses GGUF v3, general.architecture=dflash, canonical llama.cpp metadata and tensor names, and canonical one-based dflash.target_layers=[41, 42, 43].

The schema is pinned to official ggml-org/llama.cpp commit 3581ba0cf591b3f772fbb002de0f70e294bc0396, incorporating PR #25784 and the separate-DSpark converter in PR #26452. The official loader at that commit recognized the artifact as dflash, instantiated all 81 tensors, and accepted the complete layout. Its exact type inventory is F32 45, F16 2, Q8_0 25, Q2_K 9, and BF16 0.

The standardized file preserves every tensor payload from the existing custom Q2_K/Q8_0 artifact byte for byte through an explicit checked name mapping. It is not a new quantization. The legacy artifacts and their stable tags remain supported and are not deprecated pending Studio runtime validation in ns4 issue #205. BF16 is intentionally excluded from this artifact and remains a separate later phase.

Reproduce and verify the standardized candidate with:

.venv/bin/python -m unittest discover -s tests -v
.venv/bin/python scripts/dflash.py \
  --sources sources \
  --manifest-dir manifest \
  --repeat-check \
  --legacy-reference DeepSeek-V4-Flash-0731-DSpark-Drafter-Q2_K-Q8_0.gguf
sha256sum -c SHA256SUMS

Dedicated build, validation, payload-comparison, schema, and official-loader manifests are published under manifest/.

Compatibility and limitations

• This is an auxiliary drafter, not a complete checkpoint.
• The custom files require the checked-in deepseek_v4_flash_dspark_draft and dspark.* dialect. The standardized file requires llama.cpp dflash support.
• Official llama.cpp model-loading acceptance is verified for the standardized file. No Studio runtime validation or universal compatibility claim is made.
• The drafter must be paired with a compatible DeepSeek V4 Flash 0731 target model. No universal quality, acceptance-rate, or speed claim is made across target quantizations or runtimes.
• This is a community conversion, not an official DeepSeek release or endorsement.

DSpark configuration

The pinned source configuration defines block size 5, target layers 40, 41, and 42, Markov rank 256, hidden size 4096, and vocabulary size 129,280. These values are embedded in the GGUF metadata and validated during the build.

License and references

The source-derived weights remain under DeepSeek's MIT terms in LICENSE. The newly authored conversion and verification scripts are under LICENSE.code. The pinned checkpoint's upstream license and the DS4/GGML attribution in THIRD_PARTY_NOTICES.md provide the applicable notices.

• Official DeepSeek V4 Flash 0731 checkpoint
• DeepSeek V4 technical report
• DeepSpec