Google DeepMind releases EmbeddingGemma 2
A 740M-parameter open (Apache 2.0) embedding model for text, code, images, video and audio
Confirmed
The takeaway
On 6 Oct 2026 Google DeepMind released EmbeddingGemma 2, an open Apache 2.0 embedding model that maps text (including code), images, video and audio into one 768-dimensional space with 740M parameters in total, small enough for phones and laptops.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 13 October 2026
Your AI and this story
- GPT-6.2 Astra128 days after its cutoff
- Claude Opus 5.598 days after its cutoff
- Gemini 4 Hydrogen189 days after its cutoff
- Grok 4.7128 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 98 days before it.
Key facts
- 740M parameters in total: a 270M text model plus optional 170M vision and 300M audio encoders; 768-d embeddings with Matryoshka truncation to 512/256/128
- Google’s numbers: MTEB code v1 78.68 (v1: 68.76), MTEB multilingual v2 61.36, MMEB v2 overall 59.01
- Apache 2.0, not gated on Hugging Face (google/embeddinggemma-2); 8K-token context
- First Google product on it (Oct 8): Google AI Edge Foresight, a local-first, fully offline meeting transcriber and note-taker for Apple Silicon Macs, released as a developer tool. It uses EmbeddingGemma 2 for its knowledge base (PDFs, Google Docs, Office files, Markdown, bookmarks) and a Gemma 4-based assistant for Q&A, and is pitched as a Granola competitor (TechCrunch, The Verge)
- Unexpected reuse (Oct 13): the independent project Embedding Gemma Transcribe (eg2.haku.sh) turned the model into an English speech-to-text system. The audio encoder is unchanged; a small trained ‘reader’ moves each slice of audio into the text space and the closest of about 229,000 English words wins. The makers report an hour of audio in 12.6 s on a base M3 MacBook Air (288× real time, Neural Engine), a 715 MB download, and a word error rate of ‘about 5% right now, against 2–3% for leading speech models’; only the reader was trained (‘under two hours on 8×H100’). Self-reported; the Mac app is free but its source is not published. 345 points on Hacker News
What happened
Google released the second EmbeddingGemma, which moves from text-only to text, image, video and audio embeddings in one shared space, under Apache 2.0.
Embedding Gemma Transcribe (Oct 13)
An independent Mac app, Embedding Gemma Transcribe, reads words straight out of the shared embedding space: “Say a word out loud and its audio lands close to the written word.” It keeps EmbeddingGemma 2’s audio encoder as it is and trains only a small reader on openly licensed English speech (“The base model was never trained for speech-to-text”). The site reports 288× real time on a base M3 MacBook Air with 8 GB (an hour in 12.6 s), against about 150× for IBM Granite Speech 5.0 TurboCTC in the makers’ own test on the same machine, and admits a word error rate of about 5%, worse than the 2–3% of leading speech models. On Hacker News (345 points) the author, posting as Shershah1024, explained that the reader is a CTC head whose output words are EmbeddingGemma’s own text embeddings, and introduced himself as Shahir: it “started as just an experiment”, is “just a couple of days work” and supports only English for now; he wrote “I will opensource the whole thing soon.” Commenters noted that the accuracy is not yet good enough for professional use and that the app’s source code is not open yet. These figures are the makers’ own and have not been reproduced.
Why it matters
A small, permissively licensed multimodal embedding model makes on-device multimodal search and RAG practical without a hosted API.
Sources
9 sources from 8 sites. Numbers match the chips in the text.
9 sources: 3 primary, 2 press, 4 reactions
Primary
- Google blog: EmbeddingGemma 2blog.google, official
- Google DeepMind blog: EmbeddingGemma 2deepmind.google, official
- Hugging Face: google/embeddinggemma-2huggingface.co, code
Press
- TechCrunch: Google releases a new local-first Granola competitortechcrunch.com, press
- The Verge: Google’s AI note-taking app transcribes your meetings completely offlinetheverge.com, press
Reactions
- Hacker News discussion (271 points)news.ycombinator.com, discussion
- Simon Willison: EmbeddingGemma 2simonwillison.net, discussion
- Embedding Gemma Transcribe: ‘We turned it into a speech-to-text model’ (independent project)eg2.haku.sh, discussion
- Hacker News: We turned EmbeddingGemma 2 into a speech-to-text model (345 points)news.ycombinator.com, discussion
Changes
- Added Embedding Gemma Transcribe, an independent speech-to-text reuse of the audio encoder (HN front page, 345 points)
- Added Google AI Edge Foresight, the offline meeting note-taker built on EmbeddingGemma 2
- Filed