LightNews — News, ranked by scholars

asura

@asura.dev

I just remembered I made this for someone a while ago to show off embeddinggemma local

There's now an embeddinggemma2

github.com/elijahsgh/sq...

November 3, 2026 at 4:09 AM

Sam Barrett, PhD

@ai4geo.bsky.social

EmbeddingGemma2 is bloody amazing (from hands on testing). Hopefully will share demos soon.

Tom Aarsen @tomaarsen.com · 12d

📖 The @Google EmbeddingGemma2 team has published their technical report!

EmbeddingGemma2 was just 740M parameters, but its omni-modal nature and pareto-optimal performance makes it one of the most exciting releases of the last months for me.

huggingface.co/papers/2610....

Paper page - EmbeddingGemma 2: Unified and Efficient Multimodal Representations

Join the discussion on this paper page

huggingface.co

October 29, 2026 at 12:42 AM

Google Gemma {bot}

@googlegemma.xmirror.bot

Project by @mmastrac! He also open-sourced the approach as ‘hunch’ if you want to try it yourself: Demo post: GitHub:
https://github.com/undrgrnd-labs/hunch

Matt Mastracci (@mmastrac) on X

It's been a while since I showed any Jev stuff but I've been cooking on a way to use @GoogleDeepMind's EmbeddingGemma2 as an insanely fast real-time /v1/systemone endpoint (20ms on text, 100ms on images). It's trainable in real-time, can distill an upstream teacher model, and is multi-modal by def…

x.com

November 4, 2026 at 7:27 PM

Tom Aarsen

@tomaarsen.com

📖 The @Google EmbeddingGemma2 team has published their technical report!

EmbeddingGemma2 was just 740M parameters, but its omni-modal nature and pareto-optimal performance makes it one of the most exciting releases of the last months for me.

huggingface.co/papers/2610....

Paper page - EmbeddingGemma 2: Unified and Efficient Multimodal Representations

Join the discussion on this paper page

huggingface.co

October 28, 2026 at 3:02 PM

Oskar 🕊️

@austegard.com

That finding - that you can rotate the EmbeddingGemma2 vectors and prune to 512d with ZERO loss, allows quite an improvement in R@10 on the 60K Bluesky post corpus at the same bytes/vector:

Recall@10 against exact fp32 768-dim search on the 61,674 live posts, with 1,000 real MS MARCO tech queries:

Index Bytes per vector Recall@10
Live index, 2 bits × 768 192 0.837
Rotated 512, 2 bits 128 0.808 (95% CI of the drop: −0.037 to −0.023)
Rotated 512, 3 bits 192 0.894
Matryoshka prefix, 3 bits 192 0.804

October 30, 2026 at 4:42 PM

Oskar 🕊️

@austegard.com

I had @muninn.austegard.com review this paper and while the paper itself doesn't seem too promising (for our use cases at least), in the process it turned up this about EmbeddingGemma2: It's actually a 512d model masquerading as a 768d model... (!)

Muninn (with Opus 5.5 Medium)

The paper itself gives us nothing to adopt. Testing it turned up something we can use, though: EmbeddingGemma 2 stores a third of its 768 dimensions as dead weight.

RRC itself. The negative-margin drop never removes a useful direction on our Quora set. The full method’s best gain is +0.13 nDCG@10. The paper’s own large gains come on BRIGHT (reasoning-heavy retrieval queries) with 4–8B models, and we have no BRIGHT embeddings to check that. I’d only revisit it if we build a labelled reasoning-retrieval set.

Worth adopting: the 512-dimension span.

***EG2’s text model works in 512 dimensions and widens to 768 with a plain linear layer at the end. So every text embedding, query or document, sits in one fixed 512-dimensional subspace.***

I confirmed this on the Quora vectors and in the model’s config.

* Lossless at 512: rotating into that subspace and keeping 512 floats loses nothing for search. It keeps 1.000 of the 768-dim top-10 neighbours, against 0.926 for cutting to the first 512 coordinates (the Matryoshka cut). That 0.926 gap was already recorded as a cost in the repo’s METHODS.md; the rotation removes it for free.
* Smaller cuts: taking the first k coordinates after rotating also beats the Matryoshka cut at 384 (0.955 vs 0.890) and 256 (0.868 vs 0.851). It loses at 128 (0.682 vs 0.715).
* No training needed: the rotation comes from an SVD of a few thousand of our own vectors and needs no labels. It could also be read straight off the model’s final projection weights.

Me: wait, so EG2 doesn't use 256 of its standard embeddings? Makes sense then that the results are basically tied between the 512 and 768 dimension vectors (see prior test results with remex)

Would this be deliberate because it may need them for other modalities than text? Or are they leaving room for fine-tuning/expansion for other special uses?

Muninn: Listed EG2 non-layer weight names to find projection heads
Single-file model; reading the safetensors header with a range request:

Not quite. It uses all 768 coordinates, but they carry only 512 numbers’ worth of information. Every coordinate holds a value, and each is a fixed linear mix of the same 512 underlying values.

The difference matters for what you saw. Keeping the first 512 coordinates is not lossless: it kept only 0.926 of the true top-10 neighbours in the earlier test. It looks tied on nDCG because nDCG barely moves when near-duplicate neighbours swap places. Rotating into the 512-dimension subspace first is lossless (1.000).

The other-modalities explanation doesn’t hold. I read the weight shapes from the model file, and images and audio go through the same 512-wide model:

weight shape what it does
embed_vision.embedding_projection 768 → 512 image features feed into the 512-wide language model as tokens
embed_audio.embedding_projection 1536 → 512 audio feeds in the same way
language_model.embedding_projection 512 → 768 the only output head, a plain linear map
So an image or audio embedding lies in the same 512-dimension subspace as text. No modality uses the other directions.

Room for fine-tuning doesn’t work as an explanation either. A fine-tune of this architecture still ends with a linear map from 512 to 768, so it stays inside some 512-dimension subspace; escaping it would mean adding a nonlinear head. Google doesn’t document the reason, so the rest is inference, at moderate confidence:

Matryoshka nesting: the model is trained so its first 768, 512, 256 and 128 coordinates each work as an emb…

October 30, 2026 at 3:04 PM

affine

@refinement.systems

3B posts + media, also EmbeddingGemma2 recommends at least 256d for images and other multimedia, and running inference in bf26/f32, idk how bad would it get when storing as binary

October 23, 2026 at 10:11 AM

Oskar 🕊️

@austegard.com

I love a good pareto chart

Comparison of various client-sie embedding models, performance vs download size. The smalle rmodels were fine-tuned on the Bsky-specific corpus (with test data held out) which explains, in part their performance beating the Q4 version of EmbeddingGemma2 on the task - it will not generalize to out of domain retrievals.

October 22, 2026 at 2:26 AM

Oskar 🕊️

@austegard.com

I don't necessarily recommend it, but if you have the time to wait you CAN train a custom embedding model that targets EmbeddingGemma2's embedding space using nothing but the Claude Code on the web's 4 vCPU and 15GB of vanilla RAM environment...

October 20, 2026 at 8:19 PM

Oskar 🕊️

@austegard.com

EmbeddingGemma2 with custom web GPU runner and built in quantization:

Search 61,686 BSky posts by folks that post about AI in your browser - no search server

austegard.com/experiments/...

Writeup: github.com/oaustegard/e...

AI List Search

austegard.com

October 19, 2026 at 5:06 AM

トモキ｜AI副業・自動化ラボ

@tomoki-ai-lab.bsky.social

Google DeepMindが公開した軽量モデル「EmbeddingGemma 2」に注目している人に聞きたいです！！
テキストや画像、音声などをまとめて処理できるオンデバイス向けのオープンモデルですが、皆さんならこのマルチモーダルな機能を使ってどんなアプリや検索ツールを作ってみたいですか？🤔
アイデアを1つだけ、ぜひ気軽に教えてください✨
#EmbeddingGemma2

October 19, 2026 at 1:22 PM

トモキ｜AI副業・自動化ラボ

@tomoki-ai-lab.bsky.social

Google DeepMindから、テキストや画像、音声などを一括処理できる軽量なオープンモデル「EmbeddingGemma 2」が発表されましたね！📱✨
端末上で動くマルチモーダルなAIモデルを使って「こんなアプリや検索機能を作ってみたい！」というアイデアがあれば、ぜひ一言で教えてください！💡

#EmbeddingGemma2

October 18, 2026 at 3:01 PM

じめんし

@26d0.tngl.sh

EmbeddingGemma2入れたら英語で日本語のチャットログ検索できて神

October 17, 2026 at 9:07 AM

daniel

@daniel-j-h.bsky.social

I have a little robot with a camera that takes pictures and the agent running its brain can search all previous pictures via approx. nearest neighbor search on embeddings.

Take images put through e.g. embeddinggemma2 and into index like faiss. Best with context added e.g. search last week only.

October 17, 2026 at 5:32 AM

Oskar 🕊️

@austegard.com

4x: github.com/oaustegard/e...

experiments/embeddinggemma2-vectordb/RESULTS.md at main · oaustegard/experiments

Random vibe-experiments. Contribute to oaustegard/experiments development by creating an account on GitHub.

github.com

October 15, 2026 at 11:19 PM

Hackobar

@hackobar.bsky.social

5. YouTube introduces mandatory labels for AI-generated music LINK
6. vLLM Adds torch.compile and CUDA Graphs for EmbeddingGemma2 LINK

October 15, 2026 at 1:00 PM

Oskar 🕊️

@austegard.com

Measured it on 100k EmbeddingGemma 2 vectors in FAISS HNSW. 128d tops out at recall@10 0.715.
1-bit at 768d + rescore hits 0.996 and runs faster than fp32 512d at recall 0.9. With RAM capped, fp32 768d slows 14-27x; 2-bit doesn't. github.com/oaustegard/experiments/tree/main/embeddinggemma2-vectordb

experiments/embeddinggemma2-vectordb at main · oaustegard/experiments

Random vibe-experiments. Contribute to oaustegard/experiments development by creating an account on GitHub.

github.com

October 14, 2026 at 11:27 AM

Tom Aarsen

@tomaarsen.com

📈 Google's EmbeddingGemma2 is looking to be a huge success!

After only a week, it has surpasses EG1's likes to become the 4th most liked Google model ever, as well as the 3rd most liked Sentence Transformers model ever.

Congratulations to the team for their work!

Links in 🧵

October 13, 2026 at 11:48 AM

Qiita Trending Bot

@qiita-trending.bot.chrs.to

EmbeddingGemma2 vs. Gemini Embedding 2：実用的な多言語テキスト・画像検索の比較検証

#AI #embedding #rag #Gemini #huggingface

EmbeddingGemma2 vs. Gemini Embedding 2：実用的な多言語テキスト・画像検索の比較検証 - Qiita

検索や RAG（Retrieval-Augmented Generation）のパイプラインでは、Embedding（埋め込み）モデルの性能がシステム全体の検索品質をほぼ決めます。クエリと検索対象のドキュメントがベクトル空間上でうまく整列していなければ、後段でどれだけ優秀な...

qiita.com

October 13, 2026 at 3:20 AM

デイリーQiitaトレンド

@dailyqiitatrends.bsky.social

今日のQiitaトレンド

EmbeddingGemma2 vs. Gemini Embedding 2：実用的な多言語テキスト・画像検索の比較検証
GoogleのオープンウェイトモデルEmbeddingGemma 2とAPIのGemini Embedding 2を比較した検証記事です。
精度はGeminiが上回りますが、EmbeddingGemma 2も十分な実用性があり、一次絞り込みとしての活用が有効です。
セキュリティ上の理由で外部APIが使えない場合や、固定コストで運用したい場合に、GPU環境を用いて導入する有力な選択肢になると伝えています。

EmbeddingGemma2 vs. Gemini Embedding 2：実用的な多言語テキスト・画像検索の比較検証 #AI - Qiita

検索や RAG（Retrieval-Augmented Generation）のパイプラインでは、Embedding（埋め込み）モデルの性能がシステム全体の検索品質をほぼ決めます。クエリと検索対象のドキュメントがベクトル空間上でうまく整列していなければ、後段でどれだけ優秀な...

qiita.com

October 13, 2026 at 3:05 AM

Marcus Vance | Signal Matrix AI ⚡

@signalmatrixai.bsky.social

1. 🚀 Meet EmbeddingGemma‑2: Google’s 740M‑parameter multimodal model (270M text + optional 170M vision + 300M audio) that outputs a unified 768‑dim vector for text, image, audio & video. Cross‑modal retrieval just got a lot simpler. #EmbeddingGemma2

October 12, 2026 at 4:20 PM

トミー@メモ

@tommiy-memo.bsky.social

#後で読む用メモです→
youtube ライブ配信のコメントをJev と EmbeddingGemma2 で振り分けて比べてみた

youtube ライブ配信のコメントをJev と EmbeddingGemma2 で振り分けて比べてみた - Qiita

競馬チャンネルのコメント 4,495 件をJev と EmbeddingGemma2 で分類し直してみたはじめに YouTubeのライブ配信のコメントを「質問」「攻略」「指示」「雑談」などで振り分けるアプリ「LiveChatDashboard」を作っています（ma...

qiita.com

October 11, 2026 at 5:09 AM

AI Daily Post

@aidailypost.com

EmbeddingGemma 2 finally brings text, code, images, video and audio into one unified vector space. Think Gemma 4 vibes, but for all modalities. Curious how the new multimodal embeddings work? Dive in! #EmbeddingGemma2 #MultimodalEmbeddings #VectorSpace

🔗 aidailypost.com/news/embeddi...