Google’s technical report on EmbeddingGemma 2, posted to arXiv on 27 October 2026, shows its 744-million-parameter open embedding model scoring below Alibaba’s Qwen3-Embedding-0.6B on the report’s two text benchmarks and above it on code. Google launched the model on 6 October under an Apache 2.0 license.
In plain words, an embedding model turns text, images, audio or video into lists of numbers so that software can search by meaning. Google’s own tables show this one ahead on code, images and video, while Qwen3-Embedding-0.6B (596M parameters) and Jina Embed v5 Omni Nano (986M) score higher on plain text, and Jina edges it on one audio benchmark.
Short explainer · EmbeddingGemma 2 brief
A vertical news Short narrated from the published headlines and opening lines of this piece and four related FourWeekMBA pieces on the model’s launch, Qwen3 in enterprise use and on-device AI.
Business Pill · Embeddings
An evergreen Business Engineer explainer on embeddings: how a machine compares meaning. It explains the general idea only and does not discuss this report.
The key insight: As we read it, Google’s own report shows a trade-off rather than a sweep: the same 744M model that scores below Qwen3-Embedding-0.6B on both text benchmarks scores above it on code and well above Jina’s model on images and video.
On the video part of the MMEB-v2 benchmark, Google reports 50.7 for EmbeddingGemma 2 (744M parameters) against 30.7 for Jina Embed v5 Omni Nano (986M); a 19.9-point gap, our arithmetic. On text benchmarks the same report puts it below Qwen3-Embedding-0.6B. Source: Google, EmbeddingGemma 2 technical report (arXiv 2610.37364), 27 Oct 2026.
The report compares EmbeddingGemma 2 with its predecessor and with rival models in a summary table (Table 6). All scores below are the report’s, and higher is better.
Multilingual text. On MTEB(Multilingual, v2), EmbeddingGemma 2 scores 61.36, against 64.34 for Qwen3-Embedding-0.6B (596M parameters), 65.52 for Jina Embed v5 Omni Nano (986M) and 61.15 for EmbeddingGemma 1 (308M).
English text. On MTEB(English, v2), it scores 68.46, against 70.47 for Qwen3-Embedding-0.6B, 71.11 for Jina Embed v5 Omni Nano and 69.67 for EmbeddingGemma 1, so the new model is 1.21 points below its predecessor, by our arithmetic.
Code. On MTEB(Code), it scores 78.68, against 75.42 for Qwen3-Embedding-0.6B and 68.76 for EmbeddingGemma 1.
Images and video. On the image part of MMEB-v2, it scores 57.28, against 31.56 for Jina Embed v5 Omni Nano and 55.43 for InterestFM UIR CAFe (894M). On the video part, it scores 50.67, against 30.73 and 35.87.
Audio. On MAEB, Jina Embed v5 Omni Nano scores 50.14 against EmbeddingGemma 2’s 49.87. On MSEB(Retrieval), EmbeddingGemma 2 scores 69.54, against 50.95 for the text-only EmbeddingGemma 1, a score the report says it computed by transcribing the audio with speech recognition.
Google’s 6 October launch post said the model “matches the strong multilingual text performance of EmbeddingGemma” while improving code by 9.92 points, from 68.76 to 78.68. The report’s table bears that out for the multilingual score and shows the English score slightly lower.
MTEB mean task scores for EmbeddingGemma 2 (744M), Qwen3-Embedding-0.6B (596M) and EmbeddingGemma 1 (308M), as reported in Google’s EmbeddingGemma 2 technical report, Table 6. Higher is better. Benchmarks are self-reported by Google; we did not run the models. Source: Google, arXiv 2610.37364, 27 Oct 2026.
They just released EmbeddingGemma 2, an open 744M multimodal embedding model.
It beats specialist models and models 2 to 12 times larger across seven benchmark suites covering text, code, images, video, documents, and audio.
Text, images, audio, video, and mixed inputs all go into one vector space. Each modality has its own encoder feeding into a shared backbone, so you can load only what you need. Text-only runs at 271M parameters, vision-language at 439M, and audio-language at 577M.
It scores 61.36 on MTEB Multilingual v2 and 68.46 on MTEB English v2. The int4 version loses about 0.4 points on MTEB Multilingual.
You can also lower the vision token budget and truncate the embedding dimensions.
For agent memory and retrieval over mixed media, this small open model is a great on-device option.
DAIR.AI posted on 28 October that Google released EmbeddingGemma 2, an open 744M multimodal embedding model, which the post says beats specialist models and models 2 to 12 times larger across seven benchmark suites, with scores of 61.36 on MTEB Multilingual v2 and 68.46 on MTEB English v2.
Where the ‘Larger Models’ Line Comes From
The report’s abstract says the model “establishes a new Pareto frontier, outperforming both specialist models and models two to twelve times larger”. DAIR.AI’s post on X on 28 October summarised it as beating “specialist models and models 2 to 12 times larger across seven benchmark suites”.
The tables show where that holds and where it does not. Among models under 2 billion parameters on MMEB-v2, the report lists Eddy-vl_embed_1.9B_v1 (1.93B) at 63.2 against EmbeddingGemma 2’s 59.0, so the larger Eddy model scores higher on that benchmark. On text, the two rivals in Table 6 that score higher are 596M and 986M, compared with EmbeddingGemma 2’s 744M in total and 271M for its text-only configuration.
How Small It Can Run
The report says the model can run with 271M parameters as a text-only model, 439M as a vision-language model or 577M as an audio-language model, because each type of media has its own encoder feeding a shared backbone.
Memory. Google’s launch post says that, with quantization on a Google Pixel 11 Pro, the modelneeds about 191MB of active RAM for text-only weights and about 567MB for the full multimodal model.
Compression cost. In the report’s Table 1, the per-block int4 version scores 60.96 on MTEB(Multilingual, v2), 0.40 points below full precision, and 76.85 on MTEB(Code), 1.83 points below, by our arithmetic.
Context. The launch post gives an 8K-token context window, enough, it says, for up to 5.5 minutes of audio, 29 images or 58 video frames.
Who Can Use It
Google’s launch post says the model is “released under a commercially permissive Apache 2.0 license”, and the Hugging Face model card lists the license as apache-2.0, so companies can build it into products without paying Google per query. Hugging Face’s count shows about 283,000 downloads since the repository was created on 14 September, read on 28 October. Community conversions for other runtimes, including GGUF, ONNX, LiteRT, Core ML and MLX versions, appear in a Hugging Face search for the model’s name.
Google said in the launch post that the first EmbeddingGemma had been downloaded “more than 20 million” times.
Connecting the Coverage
We covered the launch on 6 October, when Google gave the model 740 million parameters and an Apache 2.0 license, as we reported then. What’s new: the technical report adds head-to-head tables against named rivals, which the launch post did not include.
Alibaba’s Qwen3 models already sit inside enterprise stacks: Korea Investment Corporation runs its in-house AI on Qwen3, with BAAI’s BGE-M3 as its search embedding model, as we reported on 19 October. Two other pieces this week dealt with AI that runs on the device. Sentient OS launched a Mac app that runs a local model overnight, as we covered on 27 October, and Apple claimed its new M6 iMac delivers up to 5x faster on-device AI than M4, as we reported on 28 October.
On the larger models, Artificial Analysis’s new index put Claude Opus 5.5 on top at 58, as we covered on 28 October.
The Structural Read
As we read it, the report is more specific than the launch post: it names the rivals and gives their scores, including the two text benchmarks where EmbeddingGemma 2 scores lower.
The Apache 2.0 license, which Google’s launch post calls “commercially permissive”, lets companies run the model on their own hardware; the report and the post give no price comparison with paid embedding services.
The report says the model can run as a 271M-parameter text-only model, and Google’s launch post gives about 191MB of active RAM for the text-only weights on a Pixel 11 Pro with quantization.
“EmbeddingGemma 2 establishes a new Pareto frontier, outperforming both specialist models and models two to twelve times larger.”
Three Implications
FOR DEVELOPERS CHOOSING A TEXT EMBEDDER In Google’s Table 6, Qwen3-Embedding-0.6B scores 64.34 and 70.47 on the two MTEB text benchmarks, against 61.36 and 68.46 for EmbeddingGemma 2.
FOR CODE SEARCH The same table gives EmbeddingGemma 2 78.68 on MTEB(Code), against 75.42 for Qwen3-Embedding-0.6B and 68.76 for EmbeddingGemma 1.
FOR MEDIA SEARCH On MMEB-v2, the report gives 57.28 on images and 50.67 on video, against 31.56 and 30.73 for Jina Embed v5 Omni Nano.
The framework puts it as “The winning strategy in every major platform shift has followed a consistent pattern: Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, EmbeddingGemma 2 sits on the open side of that line: Google’s launch post says it is built from the same technology as Gemini Embedding models and releases it under Apache 2.0. The framework was published in March 2026 and does not discuss this model.
What Is Not Established
Every score here is from Google’s own report, and we have not run the models. The rival scores are as Google reports them in its tables; we did not check them against Alibaba’s or Jina’s own figures. Google chose which models to compare against, and the report gives no prices or speed tests against rivals. Hugging Face’s download count measures file downloads, not active users.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Google’s own report shows EmbeddingGemma 2 scoring below Alibaba’s Qwen3-Embedding-0.6B on its two text benchmarks and above it on code. In the same tables it scores above Jina Embed v5 Omni Nano on images and video and slightly below it on the MAEB audio benchmark. The model is free to use under an Apache 2.0 license and can run text-only with 271 million parameters.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Benchmark scores come from Google’s EmbeddingGemma 2 technical report (arXiv 2610.37364, 27 October 2026), Tables 1, 6 and 8; rival scores are as Google reports them. Product details come from Google’s 6 October launch post and the Hugging Face model card, with download counts read from Hugging Face on 28 October. Differences between scores are our arithmetic. We did not run the models.
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.
