Google's EmbeddingGemma 2: an open 740M embedding model for text, images, audio and video, under Apache 2.0

Google released EmbeddingGemma 2 on October 6: an open embedding model that maps text, images, audio, video and code into one shared embedding space. The weights are on Hugging Face under Apache 2.0.

THE SPECS

Per Google:

- 740M parameters in total (the Hugging Face repo counts 744M). Text-only work needs as little as 270M; the vision encoder is 170M and the audio encoder 300M. - An 8K-token context window, 4x the first EmbeddingGemma. One input can hold up to 5.5 minutes of audio, 29 images or 58 video frames. - 768-dimension embeddings by default, which can be cut to 512, 256 or 128.

WHAT GOOGLE SAYS IT SCORES

These are Google's own figures. MTEB Code rises from 68.76 to 78.68 over the first version, and Google claims leading scores among sub-1B multimodal embedders on MTEB Code and MAEB (the Massive Audio Embedding Benchmark). The full table is in the model card. We have not seen an independent run.

WHERE TO RUN IT

Weights on Hugging Face and Kaggle. Google lists support in Ollama, LM Studio, llama.cpp (GGUF), MLX, vLLM, SGLang, transformers and sentence-transformers; MediaPipe and LiteRT on device; and transformers.js with WebGPU in the browser. Model Garden on Google's Gemini Enterprise Agent Platform is listed as coming soon. The Hugging Face repo showed about 99,000 downloads on October 12.

WHAT WE DO NOT KNOW

Where it lands on the public MTEB leaderboard against other open embedders, and the full language list. Google says only that it matches the first version's multilingual text performance.