Software SushiMultilingual Relevance Reranker

Available on AWS Marketplace · Version 1.0.0

Sources, licenses and distribution scope

Packaged model: BAAI/bge-reranker-v2-m3, revision 953dc6f6f85a1b2dbfca4c34a2796e7dde08d41e. The provider distributes the weights under Apache-2.0. Original license and attribution remain bundled. The offering integrates the publicly available original model; Software Sushi does not claim authorship of the weights or additional exclusive training data. The provider's upstream pretraining data was not re-created or independently audited by Software Sushi.

Integration: Software Sushi serving source and documentation are supplied under Apache-2.0. Third-party dependencies keep their own license and copyright notices. The image retains Python distribution metadata, applicable license files, shared documentation and the pinned dependency receipt. PyTorch, Transformers, safetensors and NVIDIA CUDA components are separate dependencies; the integration's Apache license does not replace their terms.

CUDA compatibility: official NVIDIA cuda-compat-13-0 package 580.196.07-1ubuntu1, SHA-256 41fea2138cad65ffdda951c720499a0ac6d9d1f4429844cd63f174a00df9f4a0. It supplies userspace compatibility libraries, not a kernel driver. Its Debian package metadata and /usr/share/doc notices remain in the image. See the official package repository and NVIDIA compatibility requirements.

Private evaluation only:

Public metrics, hashes and source attribution are retained. Evaluation source texts stay in private diagnostic storage, outside Git, the model image, documentation bundles, buyer notebooks and video assets. These sources were not used for new training. Rights decisions for private diagnostics do not assert commercial redistribution clearance for their underlying texts.

Buyer examples and video inputs are original fictional development material. Their exposed labels do not establish independent human customer acceptance. Video narration is generated with the existing NVIDIA Magpie Jason voice on Thor; no cloned person, reference recording or customer source audio is used.