Qwen3.8-27B NVFP4/Q8 Hybrid Analytical GGUF
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.
Files
| 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.
QVIR-1 R2 runtime
The validated DFlash2 and tool-calling configuration uses the public source record below:
- GitHub repository:
CostanzoPadovano/qwen38-nvfp4-analytical-lab - release branch:
codex/qvir1-release - immutable R2 publication commit:
07ccb1f3c80d7ab9875a5f0c402526d4644bd3b6 - R2 tool-trigger patch and validation:
qvir1_r2_qwen38_tool_trigger_fix_20260826.md - portable profile record:
qwen38-qvir1-r2-ngram-mod-150k.ini
The 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.
Download
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.
Validated 150K profile
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.
Validation summary
- Static GGUF validation: passed.
- Bioinformatics agent tasks: 4/4 scored 100; public tests 17/17 and hidden tests 20/20.
- Vision projector: 334/334 tensor payloads matched the verified reference.
- Long-context safe profile: validated at 98K and 131K without speculative decoding.
- QVIR-1 R2: WSL2 health check, chat-template suite, and executable FASTQ nucleotide-counting smoke test passed.
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.
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. Qwen, Unsloth,
llama.cpp, DFlash2, and related names remain the property of their respective
authors.
Citation
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