Google 稱 EmbeddingGemma 2 在多模態嵌入基準上勝過參數量最多為其兩倍的競品

Google released EmbeddingGemma 2, an open model that converts text, images, video, audio, and code into numerical vectors so similar content can be found and compared more easily. At 740 million parameters, Google says it's the most compact model of its kind and outperforms competing models up to twice its size on multimodal embedding benchmarks.

EmbeddingGemma 2 scores 78.68 on the Massive Text Embedding Benchmark (Code), a jump of nearly 10 points over its predecessor (68.76). That puts it on par with much larger models. | Image: Google

The model runs locally without an API key. Each query takes about 20 to 70 milliseconds via WebGPU in the browser. It needs only around 191 MB of RAM and cuts local vector database storage by up to six times. For text-only tasks, a 270-million-parameter version is enough.

Paired with small open models like Gemma 4, EmbeddingGemma 2 can run offline RAG apps without sending data to external servers. The weights are available on Hugging Face and Kaggle, along with a developer guide and documentation.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.