# Third-party notices

This repository distributes quantized weights derived from the models below. It
contains no third-party code and no runtime artifacts.

| Component | Upstream source and revision | License | Relationship |
|---|---|---|---|
| Base model | [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B/tree/1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0), `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0` | Apache-2.0 | Source of the BF16 linear weights that were quantized. These weights are a quantized derivative. |
| Target checkpoint | [unsloth/Qwen3.8-27B-NVFP4](https://huggingface.co/unsloth/Qwen3.8-27B-NVFP4/tree/f0b7c9e722f5565102fff8481c99e4d86ae099c7), `f0b7c9e722f5565102fff8481c99e4d86ae099c7` | Apache-2.0 | Source of the rotated Gated DeltaNet b/a rows and of the non-linear tensors used during calibration. Required at serving time and not redistributed. |
| DFlash2 draft checkpoint | [tcclaviger/Qwen3.8-27B-DFlash2-FP8](https://huggingface.co/tcclaviger/Qwen3.8-27B-DFlash2-FP8/tree/ee0cb26a8279b7910cc28d82a8a3e15e4728d56f), `ee0cb26a8279b7910cc28d82a8a3e15e4728d56f`, derived from [z-lab/Qwen3.8-27B-DFlash2](https://huggingface.co/z-lab/Qwen3.8-27B-DFlash2) (Apache-2.0) | Apache-2.0 | Required at serving time, used unchanged and not redistributed. |
| GSQ | [IST-DASLab/GSQ](https://github.com/IST-DASLab/GSQ), [arXiv:2604.18556](https://arxiv.org/abs/2604.18556) | Apache-2.0 | We reimplemented its MSE scale search and use its tensor names. No GSQ code is included. |
| GSQ 3-bit checkpoint | [ISTA-DASLab/Qwen3.8-27B-3Bit-GSQ](https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-3Bit-GSQ) | Apache-2.0 | Source of the calibration category mixture. None of its weights are included. |
| GPTQ | Frantar et al., [arXiv:2210.17323](https://arxiv.org/abs/2210.17323) | Method | Quantization algorithm, reimplemented |
| compressed-tensors | [vllm-project/compressed-tensors](https://github.com/vllm-project/compressed-tensors) | Apache-2.0 | 3-bit `pack_to_int32` storage layout. No code is included. |
| Calibration data | 33 entries, listed in [CALIBRATION.md](CALIBRATION.md) | Apache-2.0, MIT, BSD-3-Clause, ODC-By 1.0, CC BY 4.0 | Used only to collect activation statistics. The data is not redistributed. |

## Required attributions for the calibration data

**QASC.** Tushar Khot, Peter Clark, Michal Guerquin, Peter Jansen and Ashish
Sabharwal, *QASC: A Dataset for Question Answering via Sentence Composition*,
Allen Institute for AI, [allenai/qasc](https://huggingface.co/datasets/allenai/qasc)
(revision `a34ba204eb9a33b919c10cc08f4f1c8dae5ec070`). It is licensed under
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) and provided as is,
without warranties. We reformatted its facts, questions and answer options into
calibration prompts, and no QASC text is redistributed.

**FineWeb-Edu, FineWeb-2 and FinePDFs-Edu.** This model's calibration contains
information from the following datasets by Hugging Face. They are made available
under the [ODC Attribution License 1.0](https://opendatacommons.org/licenses/by/1-0/),
and their use is also subject to
[Common Crawl's Terms of Use](https://commoncrawl.org/terms-of-use).

- [HuggingFaceFW/fineweb-edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu), revision `87f09149ef4734204d70ed1d046ddc9ca3f2b8f9`
- [HuggingFaceFW/fineweb-2](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2), revision `af9c13333eb981300149d5ca60a8e9d659b276b9`
- [HuggingFaceFW/finepdfs-edu](https://huggingface.co/datasets/HuggingFaceFW/finepdfs-edu), revision `9cfabe2127faca99b3d5c4dc6d1fcb397399ebde`

The Paiton runtime image that loads these weights is distributed separately.
Upstream model, method and dataset terms remain those of their sources. The
Apache-2.0 license text is in [LICENSE](LICENSE).