Starlinghq/embeddinggemma-2-onnx

EmbeddingGemma 2, text path, ONNX (Starling build, r1)

embeddinggemma2-text-fp32.onnx is the text path of google/embeddinggemma-2 at revision 914f7f89142e33e77833254d9c9b90c3cef7303b as one fp32 ONNX graph (opset 20, PyTorch dynamo exporter): token and per-layer embeddings, the 24 encoder layers, the final norm and the 512-to-768 projection. Outputs last_hidden_state and token_embeddings. Vision, audio and video inputs are not supported by this graph; pooling, normalization, prompts and Matryoshka truncation are applied by the .NET package. NOTICE.txt lists every modification.

It is the graph the Starlings.Models.EmbeddingGemma2 .NET package (starlinghq/starling) locks by size and SHA-256 (SHA256SUMS). Reproduction: model-pipeline/reference/embeddinggemma2/export_embeddinggemma2_onnx.py followed by release_embeddinggemma2_onnx.py in that repository.

License: Apache-2.0, as the base model (LICENSE-Apache-2.0.txt, NOTICE.txt). The base model card states: Deployments must adhere to the Gemma Prohibited Use Policy.

"Gemma" is a trademark of Google LLC. StarlingHq is not affiliated with or endorsed by Google.

Private while the Starling packages are in preview: downloads need a Hugging Face token with read access to this repository.