Modelpublic

costanzopadovano/Qwen3.8-27B-DFlash2-NVFP4-GGUF

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

3likes1.3kdownloads

Model Card

Qwen3.8-27B DFlash2 NVFP4 GGUF

This repository publishes an experimental NVFP4 GGUF conversion of the `z-lab/Qwen3.8-27B-DFlash2` speculative draft model. It is intended to draft for Qwen3.8-27B targets; it is not a standalone replacement for the 27B target model.

The source weights are pinned to revision `50307d4c4cde6860d4eee73e2547cd786fe8e8a4`. The source model.safetensors SHA-256 is 67fc76d68dc5a9415511a4f394ef744d67510cd20e93b37cc2cc7d28e4bab65c; the same payload hash is present in both public source revisions inspected on 2026-08-26.

File

File	Bytes	SHA-256	Tensor layout
`Qwen3.8-27B-DFlash2-NVFP4.gguf`	1,094,346,016	`9bd8926d7d5701cee1f673dbab9e6359a739cbbfc348cb78966e8a01df8e9863`	49 NVFP4 + 32 F32

The file is approximately 1.02 GiB, 46.8% smaller than the tested Q80 draft and about 4.3% smaller than the tested Q4K_M draft.

QVIR-1 R2 runtime

This artifact was validated with the experimental QVIR-1 R2 llama.cpp runtime:

—GitHub repository: `CostanzoPadovano/qwen38-nvfp4-analytical-lab`
—release branch: `codex/qvir1-release`
—immutable R2 publication commit: `07ccb1f3c80d7ab9875a5f0c402526d4644bd3b6`
—R2 validation and benchmark: `qvir1_r2_qwen38_tool_trigger_fix_20260826.md`

Stock llama.cpp builds may not support the DFlash2, bounded-prefill, ngram-mod, Qwen3.8 tool-trigger, and experimental NVFP4 dispatch options used by this profile.

Download

bash

hf download costanzopadovano/Qwen3.8-27B-DFlash2-NVFP4-GGUF \
  Qwen3.8-27B-DFlash2-NVFP4.gguf \
  --local-dir qwen38-dflash2-nvfp4

The paired analytical target and vision projector are available at `costanzopadovano/Qwen3.8-27B-NVFP4-Q8-Hybrid-Analytical-GGUF`.

Validated settings

The QVIR-1 R2 150K profile used:

text

--spec-type ngram-mod,draft-dflash
--spec-draft-n-max 4
--spec-draft-n-min 0
--spec-draft-p-min 0.55
--spec-draft-dflash-prefill-tail 16384
--spec-draft-type-k q8_0
--spec-draft-type-v q8_0

The complete, hardware-specific profile is linked above. It used a 153,600-token target context, Q8_0 target K/V cache, and two RTX 5060 Ti 16 GB GPUs.

Measured behavior and limitations

With a Q4KXL target at long context, the NVFP4 draft saved approximately 0.90 GiB compared with Q80 while keeping long-context decode close to the Q80 result. The experimental Blackwell MMQ dispatcher was proven active, but the available sample does not establish a repeatable kernel-only speedup.

With the analytical NVFP4/Q8 Hybrid target, acceptance and throughput depend strongly on prompt content and speculative acceptance. The R2 tool-heavy benchmark showed high sample variance. This draft should therefore be treated as a VRAM-efficient experimental artifact, not as a guaranteed acceleration.

License and attribution

The source checkpoint is Apache-2.0. This derivative GGUF is distributed under the same license. See NOTICE.md for upstream attribution. DFlash2, Qwen, llama.cpp, Z-Lab, and related names remain the property of their respective authors.

Citation

Please cite the upstream DFlash2 and DFlash work as requested by the source authors, and cite the conversion/runtime record when using this artifact:

text

Costanzo Padovano. Qwen3.8-27B NVFP4 Analytical Inference Laboratory, 2026.
https://github.com/CostanzoPadovano/qwen38-nvfp4-analytical-lab