Google Launches EmbeddingGemma 2, An Open Multimodal Embedding Model
Google launched EmbeddingGemma 2, an open embedding model small enough to run on a phone, that maps text, code, images, video and audio into one shared space. The announcement post is dated October 6, 2026.
Transcript
Google just launched an open model that runs on your phone. It handles more than text.
EmbeddingGemma two maps code and media into one shared space, built on the Gemma four architecture.
Seven hundred forty million parameters keep it small enough to run offline on a phone.
It can find a video clip from a voice memo, and handles an eight thousand token context.
Google says it leads comparable small models and reports a large gain in code search on its own benchmarks.
Google published this itself, every benchmark number is its own, and no outside reviewer checked them.
Inboxsmith builds AI receptionists that answer calls for small businesses. Please like and subscribe for more news.
Sources
Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:
- This page is Google's own same day launch post for EmbeddingGemma 2, dated Oct 06, 2026 and verified live at HTTP 200 with a snapshot captured 2026-10-06T16:03:30Z. It states that EmbeddingGemma 2 natively maps combinations of text, images, audio and video into a unified embedding space, expands beyond text to also unify code, is built on the Gemma 4 architecture, has 740 million parameters, and is released under a commercially permissive Apache 2.0 license. Because the publisher is the party shipping the product, the narrated phrasing 'Google just launched' is accurate. Editorial decision: the narration says 'code and media into one shared space' rather than enumerating text, images, audio and video, because a three or more item list in narration slows the spoken read rate and has blown the duration gate before; the full modality list is carried on the silent scene 1 card at zero narrated cost. The date is never spoken.(First party Google blog announcement of EmbeddingGemma 2, backing the launch itself, the model identity, the architecture lineage and the license)
- The page reports a modular design needing as little as 270M parameters for text only workloads with optional 170M vision and 300M audio encoders; Matryoshka Representation Learning truncation from 768 dimensions down to 512, 256 or 128 for up to 6x storage reduction; approximately 191MB active RAM for text only weights and approximately 567MB for the full multimodal model on a Google Pixel 11 Pro with quantization; an 8K token context window described as 4x larger than EmbeddingGemma 1, covering up to 5.5 minutes of audio, 29 images or 58 video frames; and a 9.92 point MTEB Code improvement from 68.76 to 78.68. Editorial decision: only two figures are spoken in the entire script, 'seven hundred forty million' in scene 2 and 'eight thousand token' in scene 3, placed in different scenes and never two in one sentence, because our transcription audit enters a repetition loop when similar numeric figures are spoken close together. All numerals are spelled out in narration because the text to speech engine garbles brand tokens glued to digits and the audit misaligns when the script spells a number the transcript writes as a numeral. Every other figure, including the benchmark decimals, megabyte figures and dimension counts, was deliberately refused in narration and placed on silent cards.(Same first party Google post, backing the on device size, memory and context figures and the self reported benchmark results)
- Provenance and limits disclosure. This post is written by Sahil Dua and Henrique Schechter Vera, both Research Engineers at Google DeepMind, and published by Google on Google's own blog about Google's own model. No independent benchmark, third party evaluation, outside reviewer, auditor or journalist appears anywhere on the page, so every performance number narrated or carded is Google's own self reported figure. The linked EmbeddingGemma 2 model card was not fetched, and the referenced Google AI Edge blog post on building on device search and RAG with LiteRT was not fetched. No YouTube transcript was obtainable for this story because the local yt-dlp build is too old, so there is zero corroborating source and this single first party page backs every narrated claim. Editorial decisions on honesty: Google's marketing language, specifically 'best-in-class', 'leading scores', 'sets a new standard' and 'most capable', was refused in our own voice and moved to cards with explicit Google attribution; the source idiom 'blew past our expectations' was refused entirely and the literal download figure was carded instead; scene 4 opens with 'Google says' so no performance claim is narrated as established fact; scene 5 states plainly that Google published this itself and that no outside reviewer checked the numbers. No criticism was invented, no motive was assigned, and nothing is stated or implied to have been independently verified. The eight letter adjective form of 'agent' was kept out of every narration field because the voice engine garbles it in all configurations; it appears only on the scene 4 card inside Google's own use case list. No timeline array is included, so the date resolver uses storyDate and keeps this same day launch out of the catalog tier.(Same first party Google post, used here as the provenance and limits record for what this story does not establish)