Downloads · 30 days
0
Banana-Bae/DeepSeek-V4-Flash-0731-Q-BQ-3.0bpw
DeepSeek-V4-Flash-0731-Q-BQ-3.0bpw is a text generation model from Banana-Bae. Use it when you need the model to write or continue text. It is set up for vllm.
Q-BQ means QTIP-BananaQuant. This is the fixed FF0731 Banana Smasher QTIP2/QTIP3 mixed-anchor pack for stock vLLM.
Downloads · 30 days
0
Access
Public
Updated Aug 6, 2026
Repo size
—
Likes
0
Public
Click a slice to open those files.
.md3.8 KB · 72%
From the Hugging Face model README
Q-BQ means QTIP-BananaQuant. This is the fixed FF0731 Banana Smasher QTIP2/QTIP3 mixed-anchor pack for stock vLLM.
| Metric | Result |
|---|---|
| Top-1 | 58,389 / 65,536 = 89.094543% |
| KLD | 0.18197090324511508 |
These are the sealed BALANCED64-v1 results for this exact fixed pack. No repair or assignment rerun was performed for publication.
This repository uses banana-smasher.bpw-accounting.v1:
publication weight bytes 106,623,252,108
canonical base-model parameters 284,334,567,511
comparison BPW 2.999937799792846799052242162608052699084192147372562734142065965737483798472331993952175189063236474
publication label 3.0bpw
Comparison/publication BPW is always:
exact artifact weight bytes × 8 / canonical base-model logical parameters
Auxiliary-inclusive and expert-plane BPW values are separate accounting scopes and are not used in this model name. Repository transport also includes tokenizer/configuration metadata, repair state, and a manifest-bound ARM64 runtime wheelhouse; those support files do not change the standardized publication denominator.
The runtime assets are immutable public GitHub Release downloads:
python -m pip install --no-deps \
https://github.com/my-other-github-account/banana-smasher/releases/download/v0.2.0-sm121-runtime/deep_gemm-2.6.1-cp312-cp312-linux_aarch64.whl \
https://github.com/my-other-github-account/banana-smasher/releases/download/v0.2.0-sm121-runtime/flashinfer_python-0.6.17-py3-none-any.whl \
'https://github.com/my-other-github-account/banana-smasher/releases/download/v0.2.0-sm121-runtime/flashinfer_jit_cache-0.6.17%2Bcu130-cp39-abi3-manylinux_2_28_aarch64.whl' \
https://github.com/my-other-github-account/banana-smasher/releases/download/v0.2.0-sm121-runtime/banana_smasher_plugin-0.2.0-cp312-cp312-linux_aarch64.whl
vllm serve Banana-Bae/DeepSeek-V4-Flash-0731-Q-BQ-3.0bpw
Target environment: stock vLLM 0.24.0, CPython 3.12, Linux ARM64, CUDA 13.0, NVIDIA GB10. The plugin is discovered through the official vllm.general_plugins entry point and exposes the ordinary OpenAI-compatible vLLM API. --no-deps preserves the exact accepted wheel set and avoids the plugin wheel's build-time local wheelhouse references.
Release and checksums: Banana Smasher v0.2.0 ARM64 runtime bundle.
pack instance u012-v5
layers 43
projections 22,016
QTIP2 assignments 11,008
QTIP3 assignments 11,008
tensor layout SHA-256 0dae88283affb718f7b9cd7d6b2f9bd11016fb9b792ecf98ea96dce426ee4cc8
publication manifest SHA-256 6b9f8a8735602c5ccb1f7bcd9d3014af615117910ae53b3f071577aae534f8ae
plugin source revision f27983aec789ae78a0f27c5fe624c144f93ca9d9
The publication tree was verified against all 2,068 manifest rows after removing internal source paths from public metadata. Weight, plane, repair-state, tokenizer, configuration, and runtime-wheel bytes remain manifest-bound.
This repository is derived from deepseek-ai/DeepSeek-V4-Flash-0731; review the upstream model repository and its terms before use. Third-party runtime wheels retain their own embedded metadata and terms. No license is granted for original Banana Smasher material by this model repository.