# Third-party notices

SOYO is integration software. It does not claim ownership of the models, methods, or harnesses listed below.

## FlashVID

SOYO's optional compressed mode calls the upstream FlashVID package; it does not reimplement or vendor FlashVID.

- Project: **FlashVID: Efficient Video Large Language Models via Training-free Tree-based Spatiotemporal Token Merging**
- Authors: Ziyang Fan, Keyu Chen, Ruilong Xing, Yulin Li, Li Jiang, and Zhuotao Tian
- Repository: <https://github.com/Fanziyang-v/FlashVID>
- Pinned revision: `983cce6e30d7a8012442bfc7557d3afa61b3572d`
- Upstream package version: `0.1.0`
- License: MIT
- Upstream copyright: Copyright (c) 2026 Turbo

The upstream MIT license text is available at <https://github.com/Fanziyang-v/FlashVID/blob/983cce6e30d7a8012442bfc7557d3afa61b3572d/LICENSE> and remains in effect for FlashVID itself. Installing SOYO does not relicense FlashVID under SOYO's copyright.

## Qwen2.5-VL-7B-Instruct

SOYO loads a user-provided Qwen2.5-VL-7B-Instruct checkpoint. Model weights are not included in the SOYO package.

- Model: <https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct>
- License declared by the model card: Apache-2.0

Users remain responsible for complying with the model license and any applicable usage terms.

## DeepSeek Harness

SOYO uses the public third-party plugin interfaces of DeepSeek Harness and Cordis. DSH source is not included in SOYO.

- Repository: <https://github.com/deepseek-ai/deepseek-harness>
- Tested release: `0.1.0-rc.6`
- License declared by the upstream packages: MIT

## LongVideoBench

SOYO's benchmark records use LongVideoBench validation metadata and its multiple-choice scorer. Benchmark videos are not included in the SOYO package or repository.

- Project and code: <https://github.com/longvideobench/LongVideoBench>
- Official metadata: <https://huggingface.co/datasets/longvideobench/LongVideoBench-Meta>
- Official dataset: <https://huggingface.co/datasets/longvideobench/LongVideoBench>
- Dataset terms declared upstream: CC-BY-NC-SA 4.0
- Selective-acquisition mirror used for the V0.1 record: <https://huggingface.co/datasets/Jialuo21/LongVideoBench>
- Pinned mirror revision: `0c30edb990835a4ffd48059cd84d0680c62887a3`

The mirror's validation filenames and annotations match the pinned official metadata, and SOYO records every selected mirror LFS SHA-256. The mirror provides no provenance statement or official-tar hashes, so SOYO does not claim that its MP4 bytes are identical to the official multipart archive. Users acquiring LongVideoBench remain responsible for the upstream dataset terms.