awkeng/embeddinggemma-2-rad

EmbeddingGemma 2 — radiance container (bf16)

google/embeddinggemma-2 packed, unchanged in precision, into a single .rad container for the radiance inference engine. Every weight stays bf16 as the checkpoint ships it. This is an unofficial conversion; the model is Google DeepMind's, released under Apache 2.0 (see LICENSE and the original model card).

File	embeddinggemma2-bf16.rad — 1,502,429,184 bytes (1.39 GiB of weights, 1257 tensors)
SHA-256	d46988383e41de05df1ff69f491b1a6351a2ab69d26771943ca150768115bdae
Source	google/embeddinggemma-2 at revision 914f7f89142e33e77833254d9c9b90c3cef7303b (model.safetensors SHA-256 197a32965d4b1105faf060417baa899e193fb73cd401f42ec9295234d5553d79)
Output	768-dimensional pooled vectors (Matryoshka: dimensions may truncate) for text, images, video and audio
Engine	radiance 1.2.4 or later (embedding support); made and tested with 1.3.0

Serve

The model runs on the CPU (about 0.25–1.5 GiB of RAM) or on one GPU:

radiance --model embeddinggemma2-bf16.rad --host 0.0.0.0 --port 8000 --api-key "$KEY"

curl -s http://localhost:8000/v1/embeddings -H "Authorization: Bearer $KEY" \
  -H 'Content-Type: application/json' \
  -d '{"input":"free space path loss at 6 GHz","prompt_name":"query","dimensions":256}'

How this file was made

Converted with rad-convert from the official radiance image docker.io/stilldeadcode/radiance:1.3.0@sha256:641f2d8dc30f26dcd42673916a480bb5d1a71b28de8cd5e7fb77b47f77251b8c, with no quantisation recipe, so every weight keeps the checkpoint's encoding:

rad-convert ./embeddinggemma-2 --tokenizer ./embeddinggemma-2/tokenizer.json \
  -o embeddinggemma2-bf16.rad

./embeddinggemma-2 is the source repository at the revision above. The conversion is deterministic: running it again produces a byte-identical file with the SHA-256 above.

License

Apache License 2.0, as the original model. Copyright for the model belongs to Google DeepMind.