Google releases EmbeddingGemma 2, a 740M-parameter open multimodal embedding model scoring 78.68 on MTEB Code

Google has released EmbeddingGemma 2, an open-weights embedding model with 740 million parameters that maps text, images, video, audio, and code into numerical vectors. Google claims it is the most compact model of its kind and that it outperforms competing models up to twice its size on multimodal embedding benchmarks.

Key specifications

• Parameters: 740 million (multimodal version); a 270-million-parameter version is available for text-only tasks
• Modalities: text, images, video, audio, code
• MTEB (Code) score: 78.68, versus 68.76 for the predecessor, a gain of nearly 10 points
• Memory footprint: around 191 MB of RAM, according to Google
• Browser latency: roughly 20 to 70 milliseconds per query via WebGPU, according to Google
• Storage: Google says local vector database storage drops by up to six times
• Availability: weights on Hugging Face and Kaggle, with a developer guide and documentation
• Pricing: no API pricing; the model is distributed as open weights and runs locally without an API key
• Context window, embedding dimensions, license terms, and training cutoff: not disclosed in the available information

Benchmark results

On the Massive Text Embedding Benchmark (Code), EmbeddingGemma 2 scores 78.68. Google says that puts it on par with much larger models. The comparison models, their parameter counts, and scores on the multimodal benchmarks behind the "twice its size" claim were not itemized. The claim comes from Google and has not been independently verified.

Local and offline deployment

The model is built to run on-device. Google says a query takes about 20 to 70 milliseconds in the browser through WebGPU, with a memory requirement of about 191 MB. The company pairs it with small open models such as Gemma 4 to build retrieval-augmented generation (RAG) apps that run offline, with no data sent to external servers.

The 270-million-parameter variant covers text-only workloads, where the multimodal capacity of the larger model is not needed.

What this means

Embedding models are a quiet bottleneck in RAG pipelines. They decide what gets retrieved, and most strong ones sit behind APIs or need server-class hardware. A model of 740M parameters, or 270M for text, with a roughly 191 MB runtime footprint moves retrieval into the browser and onto edge devices, which matters for privacy-sensitive and offline applications.

The nearly 10-point gain on MTEB (Code) over the predecessor is large for a single generation. Other labs' models will still be the reference for pure text retrieval, and the headline claim of beating models twice its size rests on Google's own comparisons. Developers should test it on their own data, especially for cross-modal retrieval, before replacing an existing embedding stack.

The open weights and the single-vector-space approach to mixed media also lower the cost of building search across documents, screenshots, audio, and video without a separate model per modality. Missing details, notably license terms, context length, and embedding dimensionality, will determine how easily teams can adopt it commercially.