EmbeddingGemma 2 ships at 740M params, runs in 191MB RAM and hits 620 embeddings/sec on-device
aronchick · x · 2026-10-11
Google DeepMind released EmbeddingGemma 2: open weights under Apache 2.0, 740M parameters, mapping text, code, images, video, and audio into one shared vector space, with the text model running in about 191MB of RAM on a phone.
Alex (Alextramemory) then got it running fully on Apple's Neural Engine, sorting 10,000 posts into topics live at roughly 620 per second on a MacBook — entirely on-device, no cloud API.
The author argues embeddings unlock the rest of ML, and moving them off cloud-only APIs makes many new scenarios possible; orchestration is the next thing to watch.
Related event: Google Open-Sources EmbeddingGemma 2, a 740M Multimodal Embedding Model(3 posts)→
More from Models
- Meta says Muse Wonder beats Muse Spark 1.3's trend on unfamiliar tasks — AIatMeta · 2026-10-13
- Muse Wonder models now live on Artificial Analysis leaderboards — ArtificialAnlys · 2026-10-13
- AA Index: Muse Wonder weak on knowledge but with very low hallucination rates — ArtificialAnlys · 2026-10-13
- Muse Wonder Full tops CritPt and GDP.pdf, top-3 on six of ten AA evals — ArtificialAnlys · 2026-10-13
- SubQ claims 128x compute savings at 2M tokens with Subquadratic Sparse Attention — ryanallred · 2026-10-13
- Meta's chief AI officer taunts Google: 'gemini, who?' after Muse Wonder launch — alexandr_wang · 2026-10-13