Day 7 receipt
EmbeddingGemma 2
Google's 740M-parameter open-weight embedding model that maps text, code, images, audio and video into one shared embedding space for on-device search and retrieval.
Launch dayWhat they claimed, quoted word for word · 5
“Today, we’re launching EmbeddingGemma 2, expanding beyond text to unify code, images, video, and audio in a shared embedding space.”
Builders report using the shared space across modalities: matching audio to text embeddings for transcription, replacing an image model, and indexing text, images, audio and video locally. A llama.cpp image-embedding bug turned out to be a wrong request format.
“Built on the Gemma 4 architecture and released under a commercially permissive Apache 2.0 license, EmbeddingGemma 2 has 740 million parameters, making it optimal for on-device inference.”
The Hugging Face repo is ungated and tagged Apache 2.0. Users run it locally on laptop CPUs, an M3 Pro Mac and a macOS search tool; one CPU port matched the official embeddings closely using about 1.4 GB of memory.
“Using Matryoshka Representation Learning (MRL), developers can dynamically truncate output vectors from 768 dimensions down to 512, 256, or 128 dimensions.”
A Hindi fork of the text path measured small quality losses at 512 and 256 dimensions and a larger one at 128. Separately, vLLM rejects an explicit 768-dimension request, a serving bug rather than a model one.
“EmbeddingGemma 2 matches the strong multilingual text performance of EmbeddingGemma while delivering a significant 9.92-point improvement on code performance (in MTEB Code, from 68.76 to 78.68), making it well-suited for local codebase indexing, semantic code search, and coding agent retrieval.”
An independent 550-task multilingual retrieval benchmark found overall scores about level with the first version and a gain on code retrieval. Nobody has rerun MTEB Code itself. Text-only users report no gain, as the claim implies.
“We worked closely with the following partners to ensure EmbeddingGemma 2 works immediately where you build:”
Not immediately everywhere: Ollama's build needs an MLX runtime it lacks on Linux and Windows, vLLM fails at startup on consumer and DGX Spark GPUs, and the Quickstart misses a dependency. A llama.cpp slowdown was fixed.
Day 7What real users report · 14
A developer who runs a website on embedding models said it is a drop-in replacement for both the first EmbeddingGemma and an image embedding model, with better quality, less memory and faster embeddings.
Comment in a 439-point thread · Open the source
A user added image, audio and video input to their own local embedding tool and measured it on an M3 Pro laptop: about 78 short texts, 4 images or 6 thirty-second audio chunks a second.
Comment in a 439-point thread · Open the source
A user's first tests found it no better or faster than the first version for text-only retrieval or clustering, and preferred a specialist model for music audio.
Comment in a 439-point thread · Open the source
A user who ran it locally as a classifier found Google's own example request misclassified, and got a more accurate answer from another small model.
Comment in a 439-point thread · Open the source
The builder of an experimental speech-to-text app said it relies on text and audio sharing one vector space in the model, that it worked after a few days, and that accuracy is decent for everyday use but below cloud models.
Comment in a 435-point thread · Open the source
The creator of a 550-task multilingual retrieval benchmark found both versions scored about the same overall, with the new one ahead on code retrieval, and said the longer context is another reason to upgrade.
Hugging Face discussion · 11 comments · Open the source
A developer built a local semantic search layer for macOS files on the model, with an MCP interface so other AI tools can query it.
Hugging Face discussion · 4 comments · Open the source
A team that forked the text-only path for Hindi retrieval published measurements, including small losses when truncating to 512 or 256 dimensions and a larger loss at 128.
Hugging Face discussion · 1 comment · Open the source
Following the Quickstart in a fresh environment fails because the multimodal processor needs an image library that the install step leaves out. A Google team member said they had noted it.
Hugging Face discussion · 2 comments · Open the source
A developer released a CPU-only runtime for text, images, audio and video that beat PyTorch and ONNX Runtime on all 16 of their workloads on one laptop, matched the official embeddings closely, and used about 1.4 GB of memory.
Hugging Face discussion · 1 comment · Open the source
Pulling the model in Ollama on Linux fails because it requires an MLX runtime; others hit the same on Windows and a home server. An Ollama contributor said Linux and Windows support is coming.
Open as of 25 Oct · 6 comments · 16 reactions · Open the source
Serving it in vLLM fails at startup on consumer GPUs because one attention kernel asks for more shared memory than the cards have; two more users confirmed it on DGX Spark and a workstation GPU.
Open as of 25 Oct · 2 comments · Open the source
A user reported that llama.cpp's server ignored images sent for embedding, then closed the report the same day: image embeddings worked once they used the right request format.
Closed by the reporter as invalid · Open the source
With flash attention on, llama.cpp ran the model about ten times slower unless the input length was a multiple of 256 tokens; another user saw the same on two machines. It was closed as fixed.
Closed as completed · 1 comment · Open the source