A 92.8 GB single-file build of DeepSeek-V4-Flash-0731 (284B MoE, 13B active, 1M context) sized to run fully resident on one NVIDIA GB10 (128 GB unified memory) with room for very deep context — the 0731 DSpark speculative-decoding drafter is merged into the same file and auto-enables on load.

This repo hosts the latest release artifact (ds4flash-v5.gguf). Every release in this line is REAP-25 trimmed.

Runs only on the pulsar engine. The file uses custom tensor formats (IQ2_XXS_MMQ aligned pre-store, MXFP4/MXFP8 microscaling, BF16 source-format tensors, embedded drafter) that llama.cpp and other GGUF loaders will not accept.

What's new in v5 (srcfmt line)

v5 moves the artifact onto the source-format lineage: every tensor class carries the checkpoint's native numerics instead of an engine-side convenience format.

BF16 source-format tensors (445 tensors: embeddings, norms, residual path, LM-head base) — the source's BF16 stays BF16 end to end.
Serving fidelity is now gated against unquantized reference logits captured from the FP8/FP4 source on B300 hardware (--check-reference): KL(ref‖ours) at confident positions is 1e-5 .. 1e-6 with top-1 matched at every graded depth.
The runtime KV cache is NVFP4 on all surfaces (384 B rows) — accept rates measured at or above the previous e4m3 recipe at every depth.

What's in the file

The routed-expert formats were chosen by a measured-KL allocation — not a hand rule: per-tensor reconstruction error against the FP8/FP4 QAT source, Fisher-weighted, allocated under the byte budget by an exact knapsack.

Component	Treatment
Routed experts	IQ2_XXS floor (2.06 bpw) on 34 layers with MXFP4 (4.25 bpw, byte-lossless re-encode of the FP4 source) promoted on the quality-sensitive layers — the early stack (layers 0–11) plus four sensitive mid layers (22, 26, 29, 33). MXFP4 layers run the CUTLASS tensor-core type-40 W4A8 grouped GEMM
2-bit storage	IQ2_XXS_MMQ (type 43) aligned-SoA pre-store: the MMQ tensor-core tile layout is baked at quantize time — byte-identical logits to IQ2_XXS, no boot-time repack
Expert set	REAP-25 trimmed, per-layer policy: layers 3–42 keep 192 of 256 routed experts; the first three layers keep the full 256. The keep counts ship in the header (`reap.layer.keep_count`, layout `ds4-compact-v1`) — readable with the inspection snippet below
Attention, shared experts, dense	MXFP8 (E4M3) — byte-lossless re-encode of the FP8 QAT source, type-41 MXFP8_LT swizzle (390 tensors), loaded zero-copy
Embeddings, norms, residual, head	BF16 source format (type 30)
DSpark drafter	The 0731 drafter, embedded in the same GGUF; auto-detected and enabled at load
KV cache (runtime)	NVFP4 384 B rows on every surface (raw ring, compressed, drafter); MXFP4 indexer cache

Why MXFP4 for the promoted layers instead of a k-quant: MXFP4 is the checkpoint's source encoding, so it is a byte-lossless re-encode (zero reconstruction error) at 4.25 bpw. No k-quant can be both lossless and this small.

Provenance

Built from DeepSeek-V4-Flash source weights via gguf-tools/build/rebuild_collapsed.sh in the pulsar repo.

ds4flash-v5.gguf  92,769,087,904 bytes
sha256: 997098411c5082934c6f69bc05e22d16720e5eca4aeebeda1805a8274f5f2e8f

Requirements

NVIDIA GB10 (SM 12.1, 128 GB unified memory)
pulsar engine built with CUDA_ARCH=sm_120f

Download

Use the engine repo's ./download_model.sh v5, or:

hf download twaggs88/DeepSeek-V4-Flash-REAP25-DSpark-ds4-GGUF ds4flash-v5.gguf --local-dir gguf

Header inspection without downloading (40 MB range read):

curl -sL -H "Range: bytes=0-41943039" -o hdr.bin \
  "https://huggingface.co/twaggs88/DeepSeek-V4-Flash-REAP25-DSpark-ds4-GGUF/resolve/main/ds4flash-v5.gguf"
python3 gguf-tools/gguf_hdr.py hdr.bin   # from the pulsar repo