Pull to refresh
Logo
What Changed
Following Why Sign Up
Google DeepMind releases EmbeddingGemma 2, an open multimodal embedding model

Google DeepMind releases EmbeddingGemma 2, an open multimodal embedding model

New Capabilities

740M-parameter model maps five formats into one vector space and now powers an offline meeting-notes app

October 8th, 2026: Google releases Foresight offline meeting-notes app

Overview

Updated Oct 22

Google DeepMind released EmbeddingGemma 2 on October 6. The 740-million-parameter model turns text, code, images, video, and audio into lists of numbers that can be compared in one shared space. It runs on a phone or laptop, and its weights are open under the Apache 2.0 license.

Google's developer guide shows apps can load only what they need, from 270 million parameters for text and code to 740 million for all five formats. HuggingNews reported the model's benchmarks came close to OpenAI's embedding-3-large and Qwen 8B on consumer processors, but it cannot tell a video from its reversed version. Google also released Foresight, an offline Mac meeting-notes app powered by the model.

Why it matters

If on-device embeddings catch on, personal photos, voice notes, and documents become searchable without ever leaving the device.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

99%
Quality retained with quantized vectors
Qdrant's independent testing found full-size quantized vectors kept 99% of baseline retrieval quality using 30x less vector RAM.
78.68
MTEB Code score
Google reported a 9.92-point gain over EmbeddingGemma 1, from 68.76 to 78.68.
13 of 14
Video clips found in test
HuggingNews reported the model found 13 of 14 clips from one sentence, but it could not tell a video from its reversed version.

Voices

Curated perspectives — historical figures and your fellow readers.

Sign up to generate historical perspectives on this story.

Log in Sign Up

People Involved

SD
Sahil Dua
Author of the EmbeddingGemma 2 announcement
Announced the model launch
 Sundar Pichai
Sundar Pichai
CEO of Google and Alphabet
Announced EmbeddingGemma 2 publicly

Organizations Involved

Google DeepMind
Google DeepMind
AI company
Developer of EmbeddingGemma 2

Google's AI research lab, formed from the 2023 merger of Google Brain and DeepMind.
Google
Google
Technology company
Parent company; released EmbeddingGemma 2 and the Foresight app

Parent company of Google DeepMind and maker of Pixel phones.
Hugging Face
Hugging Face
AI model hosting platform
Distribution platform for model weights

The platform where EmbeddingGemma 2 weights are hosted.
Qdrant
Qdrant
Vector search company
Published independent testing of EmbeddingGemma 2

Vector database company that tested EmbeddingGemma 2's compression performance.

Timeline

5 events Latest: October 8th, 2026 · 1 month ago
Tap a bar to jump to that date

  1. Google releases Foresight offline meeting-notes app

    Latest Product Launch

    The free macOS app transcribes meetings and audio files without a network connection and runs EmbeddingGemma 2 alongside Gemma 4.

  2. Google DeepMind launches EmbeddingGemma 2

    Product Launch

    Open 740M-parameter model maps text, code, images, video, and audio into one 768-dimensional space. Weights released on Hugging Face and Kaggle under Apache 2.0.

  3. Sundar Pichai announces EmbeddingGemma 2 publicly

    Announcement

    Google CEO Sundar Pichai announced the model, calling it the company's first open, natively multimodal embedding model.

  4. Qdrant publishes independent compression testing

    Testing

    Vector-search company Qdrant tested EmbeddingGemma 2 across five text-retrieval datasets, finding quantized vectors retained 99% of baseline quality with 30x less RAM.

  5. Independent test finds video model lacks temporal direction

    Testing

    HuggingNews reported EmbeddingGemma 2 found 13 of 14 video clips from one sentence but treated a video and its reversed version as 99% similar.

Scenarios

1

EmbeddingGemma 2 becomes the standard on-device embedder

Likely Resolves by Apr 6, 2027

Discussed by: Google's developer blog and the model's benchmark results

If the model's benchmarks hold in practice, developers building privacy-sensitive apps adopt it. The shared tokenizer with Gemma 4 lowers the barrier for on-device RAG pipelines, and the Apache 2.0 license removes licensing friction. The 9.92-point code benchmark gain makes it attractive for local codebase indexing.

2

Competitors ship comparable open multimodal embedders

Possible Resolves by Oct 6, 2027

Discussed by: Industry analysts tracking the open-model race

Meta, Mistral, or another lab releases a sub-1B open multimodal embedding model to match. The Apache 2.0 license and the code benchmark gain set a high bar for open alternatives, but the open-model field has moved fast since Llama's release.

3

Adoption stays limited to flagship hardware

Unlikely Resolves by Oct 6, 2027

Discussed by: Skeptics noting the RAM and hardware requirements

The full multimodal model needs ~567MB of RAM, which limits it to flagship phones. If developers find the quality-per-parameter tradeoff insufficient for their use cases, adoption stays niche and the model's reach is confined to high-end devices.

4

Independent multimodal benchmarks challenge Google's claims

Possible Resolves by End of 2026

Discussed by: AI News, which noted Qdrant's testing only covers text retrieval

Qdrant validated text retrieval but did not verify image, video, or audio performance. If third-party multimodal benchmarks show the model underperforms its text results, developers may limit deployment to text-only workloads.

5

Temporal blindness limits video search to scene matching

Possible Resolves by Jan 31, 2027

Discussed by: HuggingNews testing

HuggingNews found EmbeddingGemma 2 treats a video and its reversed version as 99% similar. Developers who need to find the order of events may have to add timestamps or pair the embedder with another model, leaving video search to matching scenes rather than sequences.

Historical Context

2 moments from history that rhyme with this story — and how they unfolded.

January 2021

CLIP (2021)

OpenAI released CLIP, a model trained on 400 million image-text pairs to map images and text into a shared embedding space. It enabled zero-shot image classification and text-to-image retrieval without task-specific training.

Then

CLIP became the backbone for a wave of multimodal systems, including image generation models like DALL-E 2.

Now

It established contrastive learning on paired data as the standard way to build multimodal embeddings.

Why this matters now

EmbeddingGemma 2 extends CLIP's core idea — a shared embedding space for different modalities — from two modalities to five, and shrinks it to run on a phone.

February 2023

Llama (2023)

Meta released Llama, an open-weight large language model, under a permissive license. It showed open models could approach the performance of closed models like GPT-3.

Then

Llama sparked a wave of local AI development, fine-tunes, and on-device deployments.

Now

It established open-weight models as a credible alternative to API-only models, a trend that continued with Gemma and others.

Why this matters now

EmbeddingGemma 2 follows the same playbook — open weights, permissive license, on-device focus — applied to embeddings rather than generation.

Sources

(15)
+9
Google Blog Hugging Face Google Developers Blog Crypto Briefing Unite.AI Pasquale Pillitteri CellCog Times of India AI News Mixed News TechRepublic Verdict SiliconANGLE e/acc HuggingNews