Google DeepMind releases EmbeddingGemma 2 with 740 million parameters for on-device multimodal embeddings

Google DeepMind has introduced EmbeddingGemma 2, a compact embedding system comprising 740 million parameters made available under an Apache 2.0 license.¹ The release builds directly upon the foundational engineering that powers Google's Gemini Embedding family.¹ The system translates arbitrary combinations of written passages, still pictures, sound recordings, and video sequences into an integrated vector space.¹ By projecting disparate forms of media into a shared mathematical coordinate system, the architecture enables cross-modal retrieval directly within a single index.¹

The release follows the widespread uptake of the original EmbeddingGemma release, which accumulated more than 20 million downloads after launching last year.¹ Whereas early iterations targeted textual inputs, IT之家 reported that EmbeddingGemma 2 broadens these functional capabilities to span programming code, pictures, video, and acoustic content within that same vector space.³ This operational breadth allows downstream search pipelines to process both written language and multimedia without needing disconnected indexing pipelines.¹³ The substantial download history of the initial release established an existing base of developers seeking small-footprint embedding components.¹

To keep computation manageable across diverse edge platforms, EmbeddingGemma 2 relies on an explicitly segmented design.¹ Standard language tasks require only a base text engine configured with 270M parameters.¹ Systems handling visual or auditory data can attach an optional 170M vision encoder alongside a 300M audio encoder according to deployment demands.¹ The complete multimodal assembly is reported to total 740 million parameters, an amount that matches the combined sum of the individual text, visual, and acoustic building blocks.¹

Vector storage demands are further reduced through the integration of Matryoshka Representation Learning within the model.¹ This methodology allows operators to slice full 768-dimensional output embeddings down to narrower lengths without retraining the weights.¹ Depending on downstream latency targets and storage limits, outputs can be restricted to 512 dimensions, 256 dimensions, or 128 dimensions.¹ Smaller dimensional cuts provide substantial bandwidth and memory savings while preserving the directional geometry of the higher-dimensional vectors.¹

In addition to dimensional flexibility, EmbeddingGemma 2 expands contextual ingestion boundaries for lengthy content.¹ The framework features an 8K token context window for reading and encoding lengthy inputs.¹ This capacity represents a fourfold expansion over the context limits present in EmbeddingGemma 1.¹ The enlarged context span allows systems to process extended code files or multi-page documents in one pass without premature truncation.¹

Because the model is targeted for local execution, its active memory consumption has been calibrated for mobile processors.¹ When evaluated under quantization on a Google Pixel 11 Pro, the model's text weights demand approximately 191MB of active operational RAM.¹ Activating the full multimodal suite on the same Google Pixel 11 Pro hardware requires about 567MB of active RAM.¹ These hardware requirements ensure that even the fully loaded multimodal configuration can run alongside user tasks on handheld devices.¹

System co-location benefits from shared algorithmic components across Google's lightweight model line.² EmbeddingGemma 2 shares its text tokenizer directly with Gemma 4.² It also shares an identical audio encoder with Gemma 4.² Running the two models side by side on one machine lowers overall memory consumption by deduplicating these shared assets.²

Evaluations across standard benchmark suites demonstrate measurable accuracy improvements across technical and multilingual workloads.¹³ On the MTEB Code evaluation suite, EmbeddingGemma 2 achieved a score of 78.68, climbing from the 68.76 recorded by its direct predecessor.¹ This progression reflects an overall margin of improvement of 9.92 points over the previous generation.¹ IT之家 noted that across sub-1B parameter multimodal embedding benchmarks including MTEB Code and MAEB, EmbeddingGemma 2 delivers leading marks that equal or exceed those of larger models.³

Distribution channels have been opened across public open-weights hubs to facilitate adoption.¹² Engineers can retrieve model checkpoints directly via Hugging Face as well as Kaggle.² Google also plans to place the model into the Gemini Enterprise Agent Platform Model Garden for enterprise environments.² The Apache 2.0 licensing terms permit broad commercial integration and adaptation without proprietary restrictions.¹

To support widespread serving, the model integrates with an extensive selection of open-source deployment engines.² Implementations can be loaded through transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, and LMStudio.² For managing generated vector indexes, Qdrant is officially supported as a vector storage destination.² These serving pathways allow teams to deploy the model on both local Apple hardware via MLX and distributed backends using vLLM or SGLang.²

Edge accessibility extends into web browsers without requiring dedicated backend infrastructure.² Developers can run client-side applications inside the browser using transformers.js or through WebGPU acceleration.² For engineers looking to customize representations on specific datasets, Unsloth provides guidance for fine-tuning EmbeddingGemma 2.² Combined with flexible serving runtimes, these tools allow the 740M parameter model to be adapted and executed entirely on client endpoints.¹²