YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

EmbeddingGemma 2 β€” int8 ONNX (mirror)

Mirror of onnx-community/embeddinggemma-2-ONNX at revision daa72c51243991dfcaf9f9137d2c573d8f7790c0.

Quantized (dynamic int8, QuantType.QInt8) lanes:

  • onnx/model_quantized.onnx + _data β€” text tower (314 MB)
  • onnx/vision_encoder_quantized.onnx + _data β€” vision tower (~195 MB)
  • onnx/audio_encoder_quantized.onnx + _data β€” audio tower (~340 MB)

The tokenizer and config files are the source repo's; the graphs are byte-identical to the source revision, with sha256 in manifest.json.

Measured (2026-11-14, 1,500 long Wikipedia articles): text int8 mean 1-cos drift vs fp32 8.5e-4 (p99 2.9e-2); retrieval equal at 768/512 on the 50-query set; CPU ~0.6 docs/s, CUDA 38.6 docs/s vs fp32 38.4. See okfgraph/docs/bench-eg2-precision-width.md for the full matrix.

Upstream: EmbeddingGemma 2 by Google, Apache-2.0. This is an unmodified mirror of the onnx-community conversion; report issues against the upstream repo.

Downloads last month
30
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support