Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

EmbeddingGemma is our new best-in-class open embedding model designed for on-device AI. 📱 At just 308M parameters, it delivers state-of-the-art performance while being small and efficient enough to run anywhere - even without an internet connection.

Google DeepMind

1,532,301 subscribers

584,819 views • 1 year ago •via X (Twitter)

Health & Wellness Science & Technology Education

Anya Rossi• Live Now

Private livecam show

33 Comments

Google DeepMind1 year ago

🏆 Highest ranking on the MTEB benchmark - the gold standard for text embedding evaluation 🌐 Trained across 100+ languages 🛠️ Ready to go with @huggingface, @llama_index, @langchain and more. Here's how developers can get started with EmbeddingGemma →

JK1 year ago

The real breakthrough here is the "without an internet connection" part. On device processing is the key to making AI truly ubiquitous and private. The challenge will be maintaining this level of performance as the model is constrained by ever smaller hardware footprints.

Philip Kiely1 year ago

Awesome performance and multilingual capabilities in 300M parameters. We have day zero support on Baseten for high-throughput, low-latency deployments:

Prudent AI1 year ago

Google needs better marketing team I guess. @demishassabis why don't you guys make any noise?

Alex Veremeyenko1 year ago

embedding models getting smaller is a huge win for edge ai. excited to see how developers leverage this offline capability for privacy and efficiency.

JAKARIYA1 year ago

Love this direction. True intelligence won’t come just from more data or bigger models, but from systems that can exist, adapt, and want to survive in their environment. On-device models feel like the first step toward that.

GraphAI1 year ago

Efficiency at this size isn’t just optimization, it’s what makes contextual AI deployable anywhere

Shubham Saboo1 year ago

It's amazing. Thank you!

Himanshu Kumar1 year ago

Smaller models, bigger possibilities. Offline accessibility is a game changer for AI's future. This empowers edge computing in exciting new ways.

Tam Nguyen1 year ago

Deepmind is cooking

Send it Highor1 year ago

@demishassabis u mad about smol parameters bro? chad move = holding $TROLL 😈

Aryan Agarwal1 year ago

when are u making a better battery life in the pixel by the way please fix that shit 😔😔😔😔😔

Vastkind1 year ago

What kinds of apps could flourish once intelligence runs offline as easily as online?

Alexa | Indie hacker1 year ago

always good to see your updates, Google 💌 following along with interest

PromptSin1 year ago

Wow, that sleek on-device interface is a game-changer for AI accessibility! 🚀

C12s1 year ago

AI that actually fits in your pocket.

Anurag Pant1 year ago

👍👍👍

Sanchit monga6 months ago

Love seeing models designed for on-device from the start. 308M params with best-in-class performance is the sweet spot for real-world edge deployment. Smaller, efficient models like this are the future. We're building the runtime layer for deploying them across iOS/Android/Mac/IoT at @RunAnywhereAI.

Mykhailo Sorochuk1 year ago

Edge AI win. Small size, big impact.

OUT_OF EARTH1 year ago

@grok i am new to ai tools, so can you tell me. What it actually is? What it actually do? What this post is? What this post talking about?

Aasi Tahir Siddique1 year ago

Amazing

Masih Moafi1 year ago

what do you mean open?! Is it free?

AI PlanetX1 year ago

Impressive model, compact yet powerful!

Caden Marlow1 year ago

@grok is there similar available model ? Can i install it on my macbook pro ? Is it already pre-trained so i can discuss technical problem and brainstorm business ideas ? what are the top 3 in device model ?

Alankar Shukla1 year ago

Is there any benchmark that represents the performance between gemini embedding model and this gemma model and the openai embedding model

CoralOS1 year ago

Impressive work. In Coral, SLM orchestration with small models has already outperformed Microsoft by 34% on the GAIA benchmark, showing how efficiency can beat scale in real tasks. ✅

Dariia Hordiiuk1 year ago

On-device power.

saumya 🛠️ truth/acc1 year ago

nice

MarginCallMike1 year ago

$oscr

Pankaj Kumar1 year ago

308M parameters is impressive! Been testing it locally on my old laptop and it's surprisingly fast without sacrificing quality.

Caden Marlow1 year ago

Did someone tried it on a macbook pro ? Is this the best embedding model ? I was looking for such LLM to work during travel (train, plane ect..) i became so dependant on Cursor, and other AI tools , would be great to have a small LLM for simple task (Brainsorm on code and ideas..)

Timur Klimov1 year ago

Hey Siri

Jupiter Coder1 year ago

On-device AI getting this powerful at just 308M params is mind-blowing 🤯 huge step forward!

Related Videos

NEW: Google releases EmbeddingGemma, a state-of-the-art multilingual embedding model perfect for on-device use cases! At only 308M params, the model can run 100% locally in your browser! 🤯 Explore your documents in an interactive 3D universe with our demo: "The Semantic Galaxy"

NEW: Google releases EmbeddingGemma, a state-of-the-art multilingual embedding model perfect for on-device use cases! At only 308M params, the model can run 100% locally in your browser! 🤯 Explore your documents in an interactive 3D universe with our demo: "The Semantic Galaxy"

Xenova

23,338 views • 1 year ago

Google just proved that bigger isn't always better. Their 308M parameter model is outperforming models 2x its size. Google just released 𝗘𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝗚𝗲𝗺𝗺𝗮, and it's proving that lightweight embedding models can punch way above their weight class. At just 308M parameters (578MB), it's the new state-of-the-art for models under 500M parameters across MTEB multilingual, English, and code benchmarks. But the really impressive part is that it ranks 8th overall on MTEB(Multilingual, v2) - that's 𝟭𝟳 𝗽𝗹𝗮𝗰𝗲𝘀 above the second-best sub-500M model, and it's delivering performance 𝗰𝗼𝗺𝗽𝗮𝗿𝗮𝗯𝗹𝗲 𝘁𝗼 𝗺𝗼𝗱𝗲𝗹𝘀 𝗻𝗲𝗮𝗿𝗹𝘆 𝗱𝗼𝘂𝗯𝗹𝗲 𝗶𝘁𝘀 𝘀𝗶𝘇𝗲. There are three key parts of their training recipe that sets it apart: 𝟭. 𝗘𝗻𝗰𝗼𝗱𝗲𝗿-𝗗𝗲𝗰𝗼𝗱𝗲𝗿 𝗜𝗻𝗶𝘁𝗶𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻 Instead of starting from a decoder-only Gemma 3 model, they first adapted it to encoder-decoder, then used just the encoder. By basing EmbeddingGemma off an LLM that already has world and language understanding, it gives it a stronger starting point. 𝟮. 𝗧𝗵𝗿𝗲𝗲-𝗟𝗼𝘀𝘀 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 They combine three different loss functions, instead of just having one: • Contrastive loss (NCE) with in-batch negatives and hardness weighting • Spread-out regularization to ensure embeddings utilize the full space (for quantization and ANN retrieval) • Embedding matching distillation from Gemini Embedding - not just learning from relevance scores, but directly aligning the embedding space with the teacher model 𝟯. 𝗠𝗼𝗱𝗲𝗹 𝗦𝗼𝘂𝗽𝗶𝗻𝗴 Rather than just averaging checkpoints from the same training run, they use optimization techniques to find multiple specialized training mixtures. Each mixture creates an "expert" model in different domains, and averaging all their parameters creates a final model that's actually better than individual models. Extras: • Matryoshka embeddings supporting 768, 512, 256, and 128 dimensions • Quantization-aware training - maintains quality even at int4 precision • 100+ languages from Gemma 3 pretraining • Exceptional performance on low-resource languages (check their XTREME-UP results) Is it the absolute best embedding model? No - Gemini Embedding still leads overall. But that's not really the point. EmbeddingGemma proves you can achieve state-of-the-art performance in a small package that's actually deployable on-device, in low-latency applications, and in resource-constrained environments. This makes good embeddings accessible for use cases that I'm seeing more and more: offline applications, privacy-sensitive deployments, and high-throughput scenarios where inference cost actually matters. Full paper: Shoutout to the EmbeddingGemma team at Google DeepMind for this awesome open source work 💙 and to Daniel Williams for helping me with this video! 🫶

Google just proved that bigger isn't always better. Their 308M parameter model is outperforming models 2x its size. Google just released 𝗘𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝗚𝗲𝗺𝗺𝗮, and it's proving that lightweight embedding models can punch way above their weight class. At just 308M parameters (578MB), it's the new state-of-the-art for models under 500M parameters across MTEB multilingual, English, and code benchmarks. But the really impressive part is that it ranks 8th overall on MTEB(Multilingual, v2) - that's 𝟭𝟳 𝗽𝗹𝗮𝗰𝗲𝘀 above the second-best sub-500M model, and it's delivering performance 𝗰𝗼𝗺𝗽𝗮𝗿𝗮𝗯𝗹𝗲 𝘁𝗼 𝗺𝗼𝗱𝗲𝗹𝘀 𝗻𝗲𝗮𝗿𝗹𝘆 𝗱𝗼𝘂𝗯𝗹𝗲 𝗶𝘁𝘀 𝘀𝗶𝘇𝗲. There are three key parts of their training recipe that sets it apart: 𝟭. 𝗘𝗻𝗰𝗼𝗱𝗲𝗿-𝗗𝗲𝗰𝗼𝗱𝗲𝗿 𝗜𝗻𝗶𝘁𝗶𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻 Instead of starting from a decoder-only Gemma 3 model, they first adapted it to encoder-decoder, then used just the encoder. By basing EmbeddingGemma off an LLM that already has world and language understanding, it gives it a stronger starting point. 𝟮. 𝗧𝗵𝗿𝗲𝗲-𝗟𝗼𝘀𝘀 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 They combine three different loss functions, instead of just having one: • Contrastive loss (NCE) with in-batch negatives and hardness weighting • Spread-out regularization to ensure embeddings utilize the full space (for quantization and ANN retrieval) • Embedding matching distillation from Gemini Embedding - not just learning from relevance scores, but directly aligning the embedding space with the teacher model 𝟯. 𝗠𝗼𝗱𝗲𝗹 𝗦𝗼𝘂𝗽𝗶𝗻𝗴 Rather than just averaging checkpoints from the same training run, they use optimization techniques to find multiple specialized training mixtures. Each mixture creates an "expert" model in different domains, and averaging all their parameters creates a final model that's actually better than individual models. Extras: • Matryoshka embeddings supporting 768, 512, 256, and 128 dimensions • Quantization-aware training - maintains quality even at int4 precision • 100+ languages from Gemma 3 pretraining • Exceptional performance on low-resource languages (check their XTREME-UP results) Is it the absolute best embedding model? No - Gemini Embedding still leads overall. But that's not really the point. EmbeddingGemma proves you can achieve state-of-the-art performance in a small package that's actually deployable on-device, in low-latency applications, and in resource-constrained environments. This makes good embeddings accessible for use cases that I'm seeing more and more: offline applications, privacy-sensitive deployments, and high-throughput scenarios where inference cost actually matters. Full paper: Shoutout to the EmbeddingGemma team at Google DeepMind for this awesome open source work 💙 and to Daniel Williams for helping me with this video! 🫶

Victoria Slocum

21,610 views • 10 months ago

At the ElevenLabs Summit in Warsaw, we previewed on-device Text to Speech - a new model architecture that delivers human-level quality on limited hardware without an internet connection.

At the ElevenLabs Summit in Warsaw, we previewed on-device Text to Speech - a new model architecture that delivers human-level quality on limited hardware without an internet connection.

ElevenLabs

35,665 views • 3 months ago

.SnowflakeDB is thrilled to announce #SnowflakeArctic: A state-of-the-art large language model uniquely designed to be the most open, enterprise-grade LLM on the market. This is a big step forward for open source LLMs. And it’s a big moment for Snowflake in our #AI journey as we continue to build best-in-class enterprise-grade products for our customers. The era of enterprise AI is here. 🚀

.SnowflakeDB is thrilled to announce #SnowflakeArctic: A state-of-the-art large language model uniquely designed to be the most open, enterprise-grade LLM on the market. This is a big step forward for open source LLMs. And it’s a big moment for Snowflake in our #AI journey as we continue to build best-in-class enterprise-grade products for our customers. The era of enterprise AI is here. 🚀

sridhar

243,833 views • 2 years ago

Announcing the most powerful, intelligent, and efficient processor in its class for Windows: #Snapdragon X Elite. 💪💯 It ushers in a new era of intelligent computing with best-in-class CPU performance, leading on-device #AI, and exceptional battery life. #SnapdragonSummit

Announcing the most powerful, intelligent, and efficient processor in its class for Windows: #Snapdragon X Elite. 💪💯 It ushers in a new era of intelligent computing with best-in-class CPU performance, leading on-device #AI, and exceptional battery life. #SnapdragonSummit

Snapdragon

139,407 views • 2 years ago

Introducing ✨Tiny Aya✨, a family of massively multilingual small language models built to run where people actually are. Tiny Aya delivers strong multilingual performance in 70+ global languages in a 3.35B parameter model, efficient enough to run locally, even on a phone.

Introducing ✨Tiny Aya✨, a family of massively multilingual small language models built to run where people actually are. Tiny Aya delivers strong multilingual performance in 70+ global languages in a 3.35B parameter model, efficient enough to run locally, even on a phone.

Cohere Labs

195,494 views • 7 months ago

We’re bringing powerful AI directly onto robots with Gemini Robotics On-Device. 🤖 It’s our first vision-language-action model to help make robots faster, highly efficient, and adaptable to new tasks and environments - without needing a constant internet connection. 🧵

We’re bringing powerful AI directly onto robots with Gemini Robotics On-Device. 🤖 It’s our first vision-language-action model to help make robots faster, highly efficient, and adaptable to new tasks and environments - without needing a constant internet connection. 🧵

Google DeepMind

820,024 views • 1 year ago

Google just released Muse: a text to image model achieving state of the art performance. It is significantly more efficient than diffusion models. Muse also allows for inpainting, outpainting, and mask-free editing without finetuning or inverting the model.

Google just released Muse: a text to image model achieving state of the art performance. It is significantly more efficient than diffusion models. Muse also allows for inpainting, outpainting, and mask-free editing without finetuning or inverting the model.

bleedingedge.ai

183,750 views • 3 years ago

New AI Package with DeepSeek AI Model TesseractAI continues to expand and improve our Custom AI Agent. You will now be able to add AI packages to your agent in addition to custom information. 🌟 This new feature is being developed using #DeepSeek AI model. In particular, DeepSeek R1 is designed for greater efficiency and more goal-oriented tasks, requiring fewer computational resources while still delivering powerful results. Additionally, we’re experimenting with DeepSeek’s technology for our existing Agent LLM to further enhance performance. Our goal is to make the Custom AI Agent even more efficient and effective for all your needs.

New AI Package with DeepSeek AI Model TesseractAI continues to expand and improve our Custom AI Agent. You will now be able to add AI packages to your agent in addition to custom information. 🌟 This new feature is being developed using #DeepSeek AI model. In particular, DeepSeek R1 is designed for greater efficiency and more goal-oriented tasks, requiring fewer computational resources while still delivering powerful results. Additionally, we’re experimenting with DeepSeek’s technology for our existing Agent LLM to further enhance performance. Our goal is to make the Custom AI Agent even more efficient and effective for all your needs.

Tesseract AI

13,242 views • 1 year ago

Introducing Saaras V4 Multi-speaker Our best-in-class speech recognition model that accurately captures overlapping conversations and multi-speaker interactions, while delivering state-of-the-art performance in English across accents.

Introducing Saaras V4 Multi-speaker Our best-in-class speech recognition model that accurately captures overlapping conversations and multi-speaker interactions, while delivering state-of-the-art performance in English across accents.

Sarvam

31,294 views • 1 month ago

Introducing NexaSDK for Android (Beta) — run the latest AI models locally, 9× more energy-efficient and 2× faster, on Android devices, powered by the Qualcomm Hexagon NPU. This is the first SDK to support NPU, GPU and CPU, unlocking the full power of every Android device — for example, LFM2-1.2B achieves 85 t/s on NPU vs 37 t/s on CPU. With just 3 lines of code, you can run the latest state-of-the-art models across every task, for example: Multimodal (Vision, Audio, Text): OmniNeural-4B Embedding: EmbeddingGemma from Google ASR: Parakeet-v3 from NVIDIA OCR: PaddleOCR from Baidu Inc. Rerank: Jina-reranker from Jina AI LLM: LFM2-1.2B from And we continue to deliver Day-0 model support across our framework. Build on-device AI into your Android app today and enjoy no cloud API cost, full privacy, and offline availability. Check out our Quickstart guide and example app below to get started in minutes.

Introducing NexaSDK for Android (Beta) — run the latest AI models locally, 9× more energy-efficient and 2× faster, on Android devices, powered by the Qualcomm Hexagon NPU. This is the first SDK to support NPU, GPU and CPU, unlocking the full power of every Android device — for example, LFM2-1.2B achieves 85 t/s on NPU vs 37 t/s on CPU. With just 3 lines of code, you can run the latest state-of-the-art models across every task, for example: Multimodal (Vision, Audio, Text): OmniNeural-4B Embedding: EmbeddingGemma from Google ASR: Parakeet-v3 from NVIDIA OCR: PaddleOCR from Baidu Inc. Rerank: Jina-reranker from Jina AI LLM: LFM2-1.2B from And we continue to deliver Day-0 model support across our framework. Build on-device AI into your Android app today and enjoy no cloud API cost, full privacy, and offline availability. Check out our Quickstart guide and example app below to get started in minutes.

NEXA AI

26,091 views • 10 months ago

If your AI needs a wire to work, it isn't yours. QVAC is the local-first engine designed to run anywhere, even where the internet can't reach. Fully autonomous. Fully open source. Fully sovereign. If you can dream it, you can build it. Even in another galaxy. Start building the future of edge AI:

If your AI needs a wire to work, it isn't yours. QVAC is the local-first engine designed to run anywhere, even where the internet can't reach. Fully autonomous. Fully open source. Fully sovereign. If you can dream it, you can build it. Even in another galaxy. Start building the future of edge AI:

QVAC

7,044,421 views • 4 months ago

Alibaba just changed the AI game! Their QwQ-32B model delivers premium performance but it's completely free and open source. • Outperforms on key benchmarks • 32B parameters (vs 671B for DeepSeek) • Run locally or use web interface • Thinking mode for reasoning • Web search integration Save this video if you need powerful AI tools without the expensive subscriptions 💡 Want the SOP? DM me.

Alibaba just changed the AI game! Their QwQ-32B model delivers premium performance but it's completely free and open source. • Outperforms on key benchmarks • 32B parameters (vs 671B for DeepSeek) • Run locally or use web interface • Thinking mode for reasoning • Web search integration Save this video if you need powerful AI tools without the expensive subscriptions 💡 Want the SOP? DM me.

Julian Goldie SEO

16,923 views • 1 year ago

Introducing VisionPsy-Nano: state-of-the-art vision-language models at 460M parameters, small enough to run on the phone in your pocket. VisionPsy-Nano-460M leads every other ~0.5B vision-language model tested & compares favourably on 16 of 17 benchmarks, tops all four capability categories, and outperforms models up to 2.3x larger. Two variants, one tuned for maximum accuracy and one for real-time responsiveness. Open weights, Apache 2.0, and no cloud in the loop.

Introducing VisionPsy-Nano: state-of-the-art vision-language models at 460M parameters, small enough to run on the phone in your pocket. VisionPsy-Nano-460M leads every other ~0.5B vision-language model tested & compares favourably on 16 of 17 benchmarks, tops all four capability categories, and outperforms models up to 2.3x larger. Two variants, one tuned for maximum accuracy and one for real-time responsiveness. Open weights, Apache 2.0, and no cloud in the loop.

QVAC

15,329,496 views • 1 month ago

Not enough people are talking about NVIDIA's new Nemotron-3-Nano (4B) model! 🤯 Hybrid Mamba + Attention architecture, designed as a unified model for reasoning and non-reasoning tasks. So small and efficient, it can run 100% locally in your web browser at 75 tokens per second.

Not enough people are talking about NVIDIA's new Nemotron-3-Nano (4B) model! 🤯 Hybrid Mamba + Attention architecture, designed as a unified model for reasoning and non-reasoning tasks. So small and efficient, it can run 100% locally in your web browser at 75 tokens per second.

Xenova

50,063 views • 6 months ago

Meet Molmo: a family of open, state-of-the-art multimodal AI models. Our best model outperforms proprietary systems, using 1000x less data. Molmo doesn't just understand multimodal data—it acts on it, enabling rich interactions in both the physical and virtual worlds. Try it for yourself:

Meet Molmo: a family of open, state-of-the-art multimodal AI models. Our best model outperforms proprietary systems, using 1000x less data. Molmo doesn't just understand multimodal data—it acts on it, enabling rich interactions in both the physical and virtual worlds. Try it for yourself:

Ai2

515,700 views • 2 years ago

wow.. this new AI is shaking Hollywood LTX Studio just dropped a 13b parameters open sourced video model and.. it's now 30x faster, run on normal laptop for free check out this film trailer I created in an hour and.. new features we never see before:

wow.. this new AI is shaking Hollywood LTX Studio just dropped a 13b parameters open sourced video model and.. it's now 30x faster, run on normal laptop for free check out this film trailer I created in an hour and.. new features we never see before:

el.cine

246,707 views • 1 year ago

NEW: Alibaba just released Qwen 3.5 Small — a family of powerful multimodal models available in a range of sizes (0.8B, 2B, 4B, and 9B parameters). Perfect for on-device applications! They can even run 100% locally in your browser on WebGPU, powered by Transformers.js! 🤯

NEW: Alibaba just released Qwen 3.5 Small — a family of powerful multimodal models available in a range of sizes (0.8B, 2B, 4B, and 9B parameters). Perfect for on-device applications! They can even run 100% locally in your browser on WebGPU, powered by Transformers.js! 🤯

Xenova

102,592 views • 6 months ago

Introducing SmolVLM 256M (& 500M): The world’s smallest multimodal model. Designed for efficiency and perfect for on-device applications! 🔥 It’s so small, it can even run 100% locally in your browser on WebGPU! 🤏 Powered by Transformers.js! ⚡️ Try it out yourself! 👇

Introducing SmolVLM 256M (& 500M): The world’s smallest multimodal model. Designed for efficiency and perfect for on-device applications! 🔥 It’s so small, it can even run 100% locally in your browser on WebGPU! 🤏 Powered by Transformers.js! ⚡️ Try it out yourself! 👇

Xenova

25,302 views • 1 year ago