EmbeddingGemma 2 (AI model): news timeline & CVEs

Google’s AI note-taking app transcribes your meetings completely offline

Google launched AI Edge Foresight, a free Mac app that transcribes and summarizes meetings fully offline.

Google released Google AI Edge Foresight, a free experimental note-taking app that transcribes meetings and audio files entirely offline. It runs on macOS and is currently optimized for Apple Silicon, using the on-device EmbeddingGemma 2 model. Users can write shorthand notes that the app turns into polished notes from the transcript, query a built-in assistant, and attach local files for reference. Google says meeting audio, notes, and files never leave the computer.

The Verge · AIOct 8, 15:27 UTC · 9h agoAI industry 2 sources

Google releases a new local-first Granola competitor

Google launched offline Mac note-taker AI Edge Foresight, using on-device EmbeddingGemma 2 and Gemma 4.

Google released Google AI Edge Foresight, a Mac app positioned as a local-first competitor to AI note-taker Granola. It can run fully offline on Apple Silicon, using the 740-million-parameter EmbeddingGemma 2 model for meeting notes and a Gemma 4 assistant for questions. The app offers a split view for shorthand and AI notes, transcripts, and a knowledge base from PDFs, Docs, Office files, text, Markdown, and bookmarks. Google previously shipped an experimental local dictation tool; this release may mainly showcase Gemma’s on-device models.

TechCrunch · AIupdated Oct 8, 15:27 UTC · 9h agofirst Oct 8, 13:28 UTC · 11h agoAI industry 2 sources1

[AINews] Quasi-Riemann-Hypothesis: OpenAI publishes 722 math papers solving 90 of the top 500 open math problems; “the most significant moment” in >100 years of mathematics

OpenAI published 722 math manuscripts from an unreleased model, claiming progress on many open problems including a quasi-Riemann result.

OpenAI published 722 mathematical manuscripts in 372 families from an unreleased internal model, based on about 4,000 problems and roughly three hours of ChatGPT Pro compute per result, after consulting the Institute for Advanced Study. Commentators highlight a quasi-Riemann hypothesis result, faster-than-n-log-n integer multiplication, and a 3D elastic inverse uniqueness claim open since 1994, but the results are unverified and some researchers expect errors. The same roundup covers Mistral Large 4, a 1-trillion-parameter multimodal preview with 49 billion active parameters, priced at $1.36/$4.18 per million tokens and trained on about 3,800 Grace Blackwell GPUs in Europe, with open weights promised for the end of October.

Latent SpaceOct 7, 04:55 UTC · 1d agoAI research1

EmbeddingGemma 2

Willison praises EmbeddingGemma 2's Apache 2.0 license and warns proprietary embeddings create costly re-embedding lock-in.

Simon Willison highlighted EmbeddingGemma 2 and welcomed its Apache 2.0 license in a Hacker News comment. He argued that embedding applications often compute and store thousands or millions of vectors, so a closed, hosted-only model creates lock-in if the vendor later withdraws it. A replacement model would still require paying to recompute existing vectors. He noted that in April 2024 OpenAI offered to cover the financial cost of some users re-embedding content.

Simon WillisonOct 6, 20:37 UTC · 2d agoModel release

EmbeddingGemma 2: an open, lightweight multimodal embedding model

Google DeepMind released EmbeddingGemma 2, a 740-million-parameter open multimodal embedding model for on-device use.

Google DeepMind released EmbeddingGemma 2, a 740-million-parameter Apache 2.0 model built on Gemma 4 that maps text, code, images, video, and audio into one embedding space. It reports a 9.92-point MTEB Code gain, from 68.76 to 78.68, and can truncate vectors from 768 to 128 dimensions via Matryoshka Representation Learning. Text-only workloads can use about 270 million parameters, with optional 170-million-parameter vision and 300-million-parameter audio encoders. On a Pixel 11 Pro, quantized weights need about 191 MB RAM for text and 567 MB for the full model; context is 8K tokens. Weights are on Hugging Face and Kaggle.

Google DeepMindOct 6, 19:57 UTC · 2d agoModel release 4 sources

Google claims EmbeddingGemma 2 outperforms rival embedding models twice its size

Google released EmbeddingGemma 2, a 740-million-parameter open multimodal embedding model.

Google released EmbeddingGemma 2, an open model that turns text, images, video, audio, and code into vectors for similarity search. At 740 million parameters, Google says it is the most compact model of its kind and outperforms rivals up to twice its size on multimodal embedding benchmarks. It scores 78.68 on MTEB Code, up from 68.76, runs locally in about 20 to 70 milliseconds, and uses roughly 191 MB of RAM. A 270-million-parameter text-only variant is available, with weights on Hugging Face and Kaggle.

The DecoderOct 6, 19:47 UTC · 2d agoModel release 4 sources

Google DeepMind Releases EmbeddingGemma 2, a 740M Open Multimodal Embedding Model Built on Gemma 4

Google DeepMind released EmbeddingGemma 2, a 740M open multimodal embedding model under Apache 2.0.

Google DeepMind released EmbeddingGemma 2, a 740-million-parameter open multimodal embedding model built on Gemma 4. The model maps five input types into a single 768-dimensional embedding space. Weights are available under the Apache 2.0 license.

MarkTechPostupdated Oct 6, 19:57 UTC · 2d agofirst Oct 6, 18:36 UTC · 2d agoModel release 4 sources2

EmbeddingGemma 2: An open, lightweight multimodal embedding model

Google DeepMind released EmbeddingGemma 2, a 740M open multimodal embedding model for on-device search.

Google DeepMind released EmbeddingGemma 2, a 740-million-parameter open multimodal embedding model under Apache 2.0, built on the Gemma 4 architecture. It maps text, code, images, audio, and video into one space, with an 8K context window and Matryoshka vectors from 768 down to 128 dimensions. On MTEB Code it scores 78.68, a 9.92-point gain over EmbeddingGemma, and leads sub-1B multimodal embedders on several benchmarks. Quantized weights use about 191MB RAM for text-only and about 567MB for the full model on a Pixel 11 Pro. Weights are available on Hugging Face and Kaggle.

Hacker News · AIupdated Oct 6, 19:57 UTC · 2d agofirst Oct 6, 16:03 UTC · 2d agoModel release 4 sources HN 156↑ · 21 comments

unsloth/embeddinggemma-2-GGUF — new model trending #29 on Hugging Face

Unsloth released GGUF weights for Google DeepMind's EmbeddingGemma 2, a 740M multimodal embedding model.

Unsloth published Dynamic 3.0 GGUF weights for Google DeepMind's EmbeddingGemma 2, a 740M-parameter open multimodal embedding model under Apache 2.0. It maps text (including code), images, video, and audio into one 768-dimensional space, with optional 170M vision and 300M audio encoders around a 270M text stack. Context is 8,192 tokens, with Matryoshka truncation to 128, 256, or 512 dimensions for up to 6x less vector storage. Full-precision scores include multilingual MTEB 61.36 and code MTEB 78.68, versus 68.76 for EmbeddingGemma 1.

Hugging Face trending modelsOct 6, 10:58 UTC · 2d agoModel release 2 sources1

EmbeddingGemma 2

Timeline9 articles

Appears with

Google’s AI note-taking app transcribes your meetings completely offline

Google releases a new local-first Granola competitor

[AINews] Quasi-Riemann-Hypothesis: OpenAI publishes 722 math papers solving 90 of the top 500 open math problems; “the most significant moment” in >100 years of mathematics

EmbeddingGemma 2

EmbeddingGemma 2: an open, lightweight multimodal embedding model

Google claims EmbeddingGemma 2 outperforms rival embedding models twice its size

Google DeepMind Releases EmbeddingGemma 2, a 740M Open Multimodal Embedding Model Built on Gemma 4

EmbeddingGemma 2: An open, lightweight multimodal embedding model

unsloth/embeddinggemma-2-GGUF — new model trending #29 on Hugging Face