Downloads · 30 days
44
1% of all-time downloads
costanzopadovano/Qwen3.8-27B-NVFP4-Q8-Hybrid-Analytical-GGUF
Qwen3.8-27B-NVFP4-Q8-Hybrid-Analytical-GGUF is a image-text-to-text model from costanzopadovano. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
This repository publishes the target model used by the QVIR-1 analytical inference experiments. It is a one-pass GGUF conversion of unsloth/Qwen3.8-27B-NVFP4 at the pinned revision 16b6615af3548b88e2d8e382457bc705b004…
Downloads · 30 days
44
1% of all-time downloads
All-time downloads
3.3K
Public
Repo size
22.5 GB
Likes
2
Public
Click a slice to open those files.
.gguf22.5 GB · 100%
From the Hugging Face model README
This repository publishes the target model used by the QVIR-1 analytical
inference experiments. It is a one-pass GGUF conversion of
unsloth/Qwen3.8-27B-NVFP4
at the pinned revision
16b6615af3548b88e2d8e382457bc705b00479cf.
The conversion preserves 168 native NVFP4 matrices, converts 338 analytically sensitive matrices to Q8_0, and retains 696 tensors as F32. The result was selected for local bioinformatics, coding, long-context, and vision workloads after controlled validation; it is not an official Qwen, Unsloth, ggml-org, or Z-Lab release.
| File | Bytes | SHA-256 |
|---|---|---|
Qwen3.8-27B-NVFP4-Q8-Hybrid-Analytical-v1.gguf | 21,560,282,304 | 3d5309691e6f5467dcd7514ca8fe7443fb4b05cbd6b0d980d70572a0d9a7dc12 |
mmproj-F16.gguf | 927,607,488 | cbb841a9ee0636b2ec172f5bb8df2ea8dfeb01e90fe7c6126581d662a0b4e43e |
The target contains 1,202 tensors: 168 NVFP4, 338 Q8_0, and 696 F32. Its declared native context is 262,144 tokens. The F16 projector contains 334 tensors and passed tensor-by-tensor payload equivalence against the verified Qwen3.8 F16 projector.
The validated DFlash2 and tool-calling configuration uses the public source record below:
CostanzoPadovano/qwen38-nvfp4-analytical-labcodex/qvir1-release07ccb1f3c80d7ab9875a5f0c402526d4644bd3b6qvir1_r2_qwen38_tool_trigger_fix_20260826.mdqwen38-qvir1-r2-ngram-mod-150k.iniThe R2 runtime is experimental. Stock llama.cpp builds may not understand all
of the DFlash2, bounded-prefill, ngram-mod, Qwen3.8 tool-trigger, or NVFP4
options used by the validated profile.
hf download costanzopadovano/Qwen3.8-27B-NVFP4-Q8-Hybrid-Analytical-GGUF \
Qwen3.8-27B-NVFP4-Q8-Hybrid-Analytical-v1.gguf \
mmproj-F16.gguf \
--local-dir qwen38-hybrid
For DFlash2 speculative decoding, pair this target with
costanzopadovano/Qwen3.8-27B-DFlash2-NVFP4-GGUF.
The R2 experiment used two NVIDIA RTX 5060 Ti 16 GB GPUs, a 153,600-token
context, Q8_0 K/V cache, tensor split 0.80,1.20, batch 1024, ubatch 128,
F16 vision projection, DFlash2 NVFP4 with four-token proposals, bounded DFlash
prefill over the final 16,384 tokens, and ngram-mod.
This is a hardware-specific fit point, not a universal recommendation. The same target/draft pair may require a smaller context, smaller batch, different tensor split, or no vision projector on other systems.
The R2 benchmark preserved no-tool throughput and produced a favorable aggregate result on the tested long tool-heavy workload, but speculative acceptance was highly variable. No universal speedup is claimed.
The source checkpoint is Apache-2.0. This derivative GGUF is distributed under
the same license. See NOTICE.md for upstream attribution. Qwen, Unsloth,
llama.cpp, DFlash2, and related names remain the property of their respective
authors.
If you use the conversion, runtime record, or measurements, please cite the repository:
Costanzo Padovano. Qwen3.8-27B NVFP4 Analytical Inference Laboratory, 2026.
https://github.com/CostanzoPadovano/qwen38-nvfp4-analytical-lab