Google's EmbeddingGemma 2 drops the Gemma licence gate, and gains 0.21 on multilingual text
Google's new on-device embedding model can be downloaded without asking Google first, and the previous one still needs Google's approval. The Hugging Face record for google/embeddinggemma-2 carries the tag license:apache-2.0 with gating set to false, while the record for EmbeddingGemma 1 carries license:gemma with gating set to manual, the setting that holds a download behind a request Google approves by hand. Both were read at 23:06 UTC on October 6.

A scatter chart of mean score against model size, with EmbeddingGemma 2 marked in blue among grey dots for other models. Image: Google
EmbeddingGemma 2 was announced on October 6 by Sahil Dua and Henrique Schechter Vera, research engineers at Google DeepMind, in Google's launch post. It has 740 million parameters, a 270M text backbone with vision and audio encoders that load only when a workload needs them, and it maps text, code, images, video and audio into one 768-dimensional space.
The licence change is the part that moves
Apache 2.0 on its own is not new for a Gemma derivative, as Google's 4B speaker-labelling model from October 5 shipped under the same licence. The gate is what changes. Version 1 sits under the Gemma Terms of Use with manual approval switched on, and Google says it passed "more than 20 million downloads" that way. Version 2 is Apache 2.0 with the gate off, so a build server can pull the weights unattended.
The model card lists "License: Apache 2.0" and names Google DeepMind as the author. Weights are on Hugging Face and Kaggle, and Google says Gemini Enterprise Agent Platform Model Garden availability is "coming soon".
The gain is in code, not in multilingual text
Google ran the benchmarks itself, on the full-precision checkpoint, and published them against EmbeddingGemma 1. On MTEB code v1 the new model scores 78.68 against 68.76, a gain of 9.92 points. On MTEB multilingual v2 it scores 61.36 against 61.15, a gain of 0.21.
The image, video and audio rows carry no comparison figure at all, because EmbeddingGemma 1 handled text only. Google's numbers for the new modalities are 64.64 on MIEB lite, 57.28 and 67.84 on the two MMEB v2 image tracks, 50.67 on MMEB v2 video, 69.54 on MSEB retrieval and 49.39 on MAEB.
Google's launch post calls the model "Best-in-class for its size" and claims "leading scores among sub-1B multimodal embedders". The model card settles for "among the strongest multimodal embedding models under 1B parameters", and every comparison figure Google published there is against EmbeddingGemma 1 and nothing else. Rival models appear only in the launch post's scatter charts, where Google plots them by size rather than tabulating the scores.
Shrinking the vector costs more than the headline figure
Matryoshka Representation Learning lets a 768-dimensional vector be cut to 512, 256 or 128, which Google sells as "up to 6x storage reduction for local vector databases and memory usage". At 128 dimensions the MMEB v2 overall score falls from 59.01 to 45.65, and MSEB retrieval from 69.54 to 56.71.
Google says as much on the model card: "128 dimensions degrades multimodal quality substantially and should be validated against your own workload before adoption." Quality holds up better at 256 dimensions, where the overall multimodal score is 56.24.
On hardware, Google's figure is quantised weights on a Google Pixel 11 Pro needing "as little as ~191MB active RAM for text-only weights and ~567MB for the full multimodal model". The context window is 8,192 tokens, four times EmbeddingGemma 1, shared across modalities at 25 tokens per second of audio and 280 tokens per image, which Google translates as 5.5 minutes of audio, 29 images or 58 video frames in a single pass.
Training data runs to a January 2025 cutoff across more than 140 languages. Every score above is Google measuring Google, on a checkpoint Google published itself.