Google DeepMind said on 6 October 2026 that it is launching EmbeddingGemma 2, which its launch post describes as “expanding beyond text to unify code, images, video, and audio in a shared embedding space.” The post gives the model 740 million parameters and an Apache 2.0 license, and says the weights are on Hugging Face and Kaggle.
The model card says it “maps text (incl. code), images, video, and audio inputs—and combinations thereof—into a single, unified 768-dimensional vector space.” Everything below about quality is Google’s own account, from the launch post and the card. We haven’t run the model or checked its benchmark results independently.
Update, 28 October 2026: Google posted a technical report on EmbeddingGemma 2 to arXiv on 27 October that compares the model with named rivals, including Alibaba’s Qwen3-Embedding-0.6B. We cover what its tables show in Google Paper: EmbeddingGemma 2 Trails Qwen on Text, Not Code.
Business Pill · MEANING AS DISTANCE
A one-minute explainer of the idea behind this story: an embedding. It teaches the concept, not this story’s figures.
The key insight: On the text benchmark that carries over from the first model, the model card shows 61.36 against 61.15, and the launch post says the model “matches the strong multilingual text performance of EmbeddingGemma.” The changes the documents report are elsewhere: a 9.92-point rise on the code benchmark, and image, video and audio inputs mapped into the same 768-dimension space.
The launch post is headed “EmbeddingGemma 2 is a best-in-class open model for natively multimodal embeddings.” It calls the model “the most capable model for on-device multimodal embeddings” and carries the names of two Google DeepMind research engineers, Sahil Dua and Henrique Schechter Vera.
It says the first EmbeddingGemma was introduced last year as “a lightweight option for high-quality text embeddings,” and that the first model has had “more than 20 million downloads.”
The model card, on Google’s developer site and marked as last updated on 6 October 2026, gives 740 million parameters in total. They are a 270 million-parameter text model plus a 170 million-parameter vision encoder and a 300 million-parameter audio encoder. It lists 24 layers, a context window of 8,192 tokens and a native output of 768 dimensions.
The launch post says the model is “Built on the Gemma 4 architecture and released under a commercially permissive Apache 2.0 license.” The card says it understands 100+ languages and gives its pre-training data a cutoff date of January 2025.
EmbeddingGemma 2 scores at the full 768 dimensions and at 128 dimensions, as the model card reports them. The benchmarks use different metrics. The figures are Google’s account; we have not checked them.
The Use Cases the Post Describes
The post offers two examples: “It can help find a specific video clip from a voice memo, or search through hours of audio recordings based on a text query, all processed by a single, natively multimodal model.”
It argues for running the model on the device itself: “Generating embeddings locally helps ensure data privacy, reduces pipeline latency, and empowers developers to build cross-modal search and retrieval that works entirely offline.”
The post adds that because EmbeddingGemma 2 is built on Gemma 4 and shares its text tokenizer and audio encoder, developers can run both models together in one pipeline with a lower combined total memory footprint.
The Scores Google Reports
The card lists results at the full 768 dimensions and says “All results reported below use the full-precision checkpoint.” Against its predecessor, it reports 61.36 against 61.15 on the multilingual MTEB text benchmark, and 78.68 against 68.76 on the MTEB code benchmark.
By our subtraction, the multilingual gap is 0.21 of a point and the code gap is 9.92 points. The launch post says the model “matches the strong multilingual text performance of EmbeddingGemma while delivering a significant 9.92-point improvement on code performance.”
For the other inputs the card shows no figure for the first model: its table has a dash in that column. It reports 64.64 on the MIEB (lite) image benchmark, 57.28 on the image part of MMEB v2, 67.84 on its visual-document part and 50.67 on its video part. It reports 69.54 on the MSEB audio retrieval benchmark and 49.39 on MAEB. The card lists several metrics across these rows, including Hit@1, NDCG@5 and MRR@10, so the figures are not on one scale.
The post claims the model “sets a new standard in quality-per-parameter for sub-1B models and even outperforms some specialist models more than twice its size.” The card says it is “among the strongest multimodal embedding models under 1B parameters.” The text of the post and the card names no competing model.
What Shrinking the Vectors Costs
The model is trained with Matryoshka Representation Learning. The post says that lets developers “dynamically truncate output vectors from 768 dimensions down to 512, 256, or 128 dimensions,” for “up to 6x storage reduction.”
The card’s truncation table shows the price. At 256 dimensions the multilingual MTEB score is 60.41 and the MMEB v2 overall score is 56.24, against 61.36 and 59.01 at 768. At 128 dimensions they are 57.89 and 45.65.
By our subtraction, 128 dimensions costs 3.47 points on multilingual text and 13.36 points on MMEB v2. The card says “Quality is close to lossless down to 256 dimensions,” and that “128 dimensions degrades multimodal quality substantially and should be validated against your own workload before adoption.” It adds that “128d is best suited to text-only workloads.”
Running It on a Phone
The post says: “With quantization, on a Google Pixel 11 Pro, EmbeddingGemma 2 requires as little as ~191MB active RAM for text-only weights and ~567MB for the full multimodal model.”
The card lists the effective size by configuration: 270 million parameters for text only, 440 million for text and image, 570 million for text and audio, and 740 million for the full model. The post says the 8K context window lets the model process “up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved combinations thereof.”
The two documents report on different versions of the model. The RAM figures are for a quantized model, and the card says its benchmark results use the full-precision checkpoint.
Cautions in the Model Card
The card says to run inference in bfloat16 or float32 and “Do not use float16.” It explains: “In float16 the model returns NaN or silently degraded embeddings rather than raising an error, so the failure is easy to miss.”
For text inputs the card recommends a short task prefix. Using the right one improves quality, it says, and omitting it “still works but reduces precision.”
The card says the model is “a pre-trained embedding model” that does not undergo “post-training alignment, safety tuning, or output-level moderation,” and that developers and deployers are responsible for application-level safeguards. On languages it says “the model may not exhibit equal performance across languages.”
The card’s header reads “License: Apache 2.0.” The same card says “Deployments must adhere to the Gemma Prohibited Use Policy.” The metadata of the model’s Hugging Face repository lists the license as apache-2.0. We did not read the Gemma Prohibited Use Policy or a separate license text.
Where to Get It
The post says the weights are on Hugging Face and Kaggle, “with Gemini Enterprise Agent Platform Model Garden availability coming soon.” It lists transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama and LMStudio for serving the model, LiteRT and MediaPipe for on-device use, and transformers.js or WebGPU for the browser. It points to guidance by Unsloth for fine-tuning.
The Structural Read
Our reading of the post and the card: they make three kinds of claim. One is capability, four kinds of input in one space. One is footprint, with effective sizes from 270 million to 740 million parameters and, for a quantized model on a Google Pixel 11 Pro, 191MB to 567MB of RAM. One is economy, vectors that can be cut to a sixth of their size.
The card states several trade-offs itself. It says 128 dimensions degrades multimodal quality substantially, tells developers not to use float16, and says the model has no post-training alignment or safety tuning of its own.
The strongest wording in the post, quality per parameter and specialist models more than twice its size, comes without a named model in the text of either document.
Google DeepMind, in its EmbeddingGemma 2 launch post
“Generating embeddings locally helps ensure data privacy, reduces pipeline latency, and empowers developers to build cross-modal search and retrieval that works entirely offline.”
Three Implications
THE SIZE IS A MENU The card lists four configurations: 270 million parameters for text only, 440 million with images, 570 million with audio and 740 million for everything. The post says the vision (170M) and audio (300M) encoders are optional.
SMALLER VECTORS COST MORE ON SOME INPUTS By our subtraction from the card’s truncation table, 128 dimensions costs 13.36 points on MMEB v2 and 12.83 on MSEB retrieval, against 3.47 on multilingual text. The card says “128d is best suited to text-only workloads.”
SETTINGS CARRY WEIGHT The card says omitting the task prefix “still works but reduces precision,” and that in float16 the model returns NaN or silently degraded embeddings. Both appear in the card’s best-practices guidance.
What Is Not Established
The text of the launch post and the model card names no competing model, so we can’t say which models the quality-per-parameter claim and the “specialist models” claim refer to. The card’s only reference column is the first EmbeddingGemma, and only for the two text benchmarks.
We haven’t run the model, tested it on a phone or checked any score. The scores are Google’s account, reported for the full-precision checkpoint. We did not read the Google AI Edge blog post, the weights, the Gemma Prohibited Use Policy or other outlets’ coverage of the launch.
We did not contact Google. This piece draws no conclusion about the model or the company. Readers can see our explainer Open Weights Explained and our earlier report on the Mistral Large 4 preview. Neither involves a comparison with this model.
Business Engineer Framework
The Map of AI — The Stack Beneath the Applications
The Map of AI places more than 200 companies across nine layers of the stack, from silicon to application. It shows where model builders sit relative to the chips below them and the applications above.
Google DeepMind said on 6 October 2026 that it is launching EmbeddingGemma 2, a 740 million-parameter embedding model for text, code, images, video and audio, under an Apache 2.0 license, with weights on Hugging Face and Kaggle. The model card reports a code score of 78.68 against 68.76 for its predecessor, and shows the scores falling as the vectors shrink to 128 dimensions. Every score is Google’s own account, and we haven’t checked any of them.
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. This piece rests on Google DeepMind’s launch post of 6 October 2026, “EmbeddingGemma 2 is a best-in-class open model for natively multimodal embeddings,” and on the EmbeddingGemma 2 model card on Google’s developer site, both read in full, plus the metadata of the model’s Hugging Face repository. The benchmark scores are Google’s account. We haven’t run the model, tested it on a phone or checked any score, and we didn’t read the Google AI Edge blog post, the Gemma Prohibited Use Policy, the weights or other outlets’ coverage of the launch. We did not contact Google. Nothing here is a forecast, and nothing here is financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
