EmbeddingGemma
GoogleEmbeddingGemma is Google's 308M-parameter multilingual text embedding model, built on Gemma 3 and aimed at running on phones, laptops and tablets rather than in a datacenter. It takes up to 2K tokens of context and emits 768-dimension vectors, truncatable to 512, 256 or 128 via Matryoshka Representation Learning, and is distributed alongside a quantization-aware-trained checkpoint for low-memory deployment. The weights are downloadable but gated behind acceptance of the Gemma Terms of Use.
The Hugging Face card itself returns 401 because the repository is gated, so the license and evaluation facts are read from Google's published model card instead. Verified 2026-09-11 via the Google AI model card, the Gemma Terms of Use and the two Hub API records.
Openness
3 high confidence- weights
- open(downloadable but gated: the Hub repo requires manual acceptance of the Gemma terms)
- data
- closed(approximately 320B tokens of web documents, code and synthetic text, characterized by type only)
- code
- partial(inference and fine-tuning guides
- license
- Gemma-License(Gemma Terms of Use, use bounded)
The Gemma Terms of Use permit commercial use but reserve the right to restrict it and bind every downstream distribution to the same prohibited-use policy, which is the use_bounded rung: a real bound that almost no user meets. The gate is friction rather than a further restriction - anyone may accept and download - and the training corpus is described only by category, so nothing pulls the score above the open-weights rung.
- https://ai.google.dev/gemma/docs/embeddinggemma/model_card recorded 2026-09-11
Google's model card gives the 2K input context, 768-dimension output with MRL truncation, and a training dataset of approximately 320 billion tokens described by component type - web documents, code, synthetic - with no dataset released.
- https://ai.google.dev/gemma/terms recorded 2026-09-11
The Gemma Terms of Use grant a non-exclusive, royalty-free right to use and distribute the model, subject to the Prohibited Use Policy which must be passed on to every downstream recipient, and reserve Google's right to restrict use it believes violates those terms.
- https://huggingface.co/api/models/google/embeddinggemma-300m recorded 2026-09-11
gated: manual; private: false; cardData license gemma - the weights are downloadable after accepting the terms.
Adoption
4 high confidence2,132,311 downloads in the trailing 30 days across the two shipped checkpoints - 2,129,726 for embeddinggemma-300m and 2,585 for the QAT variant. Notable because the repository is gated: this is volume reached through a click-through license rather than an open download.
- https://huggingface.co/api/models/google/embeddinggemma-300m recorded 2026-09-11
downloads: 2129726; gated: manual; private: false.
- https://huggingface.co/api/models/google/embeddinggemma-300m-qat-q8_0-unquantized recorded 2026-09-11
downloads: 2585; gated: manual; private: false.
Capability
3 high confidenceRung 3. Genuinely strong for 308M parameters and the best of the on-device multilingual options, but the rung is about retrieval quality rather than efficiency, and 61.15 on MMTEB v2 against the anchor's 74.3 is a wide gap. The 2K context is the other narrowing: a sixteenth of Harrier-OSS's window, which rules out the long-document retrieval the frontier models are sold on. Two below the anchor on the same instrument.
- https://ai.google.dev/gemma/docs/embeddinggemma/model_card recorded 2026-09-11
Full-precision checkpoint results: MTEB (Multilingual, v2) mean 61.15 at 768d, 60.71 at 512d, 59.68 at 256d and 58.23 at 128d; MTEB (English, v2) mean 69.67 at 768d.
- https://ai.google.dev/gemma/docs/embeddinggemma recorded 2026-09-11
Overview describes a 308M-parameter multilingual embedding model based on Gemma 3, optimized for everyday devices such as phones, laptops and tablets.
Verified 2026-09-11