← back
Gemma, DeepMind's Family of Open Models — Omar Sanseviero, Google DeepMind
Original: Gemma, DeepMind's Family of Open Models — Omar Sanseviero, Google DeepMind
Takeaway
Gemma 4 makes frontier open-model intelligence runnable on phones and single consumer GPUs, with on-device agentic and multimodal use cases.
Summary
- Omar Sanseviero introduces Gemma 4 (2B–32B), Google DeepMind's most capable open model family, just 7 days after launch.
- Smallest variants run on-device on Android, iPhone and Raspberry Pi with multimodal + reasoning support; demos show piano-playing skills agent and offline coding fully in airplane mode.
- MoE variant offers very low latency; the 31B is the most intelligent and still fits in a single consumer GPU.
- Talk positions Gemma against open models on LMSYS arena and emphasizes Google's developer-friendly H100/A100 footprint chart.
gemmaon-deviceopen-models
Original description
Google DeepMind’s Gemma family is expanding. Join us for a deep dive into the latest models of the Gemma ecosystem. From vibe fine-tuning to Sovereign AI, you'll learn about the latest model capabilities, how to build high-performance applications, and how to get started with open models. Speaker info: - https://x.com/osanseviero - https://www.linkedin.com/in/omarsanseviero/ - https://github.com/osanseviero Timestamps 0:00 Introduction to the Gemma model family 0:41 Evolution from Gemma 3 to Gemma 4 1:21 Overview of the new Gemma 4 capabilities 2:31 Live demonstrations of on-device applications 3:38 LM Arena scores and performance benchmarks 5:07 Apache 2 license transition 5:27 Technical deep dive: The E2B architecture and per-layer embeddings 6:57 Multimodal understanding and multilingual support 8:43 Ecosystem growth and community adoption 10:07 Product integrations, including Android Studio 10:46 Statistics on model downloads and fine-tuning 11:27 Official Gemma variants: Shield Gemma and MedGemma 12:16 Community research and sovereign AI efforts 12:56 Real-world applications, from cancer therapy to offline tasks 14:05 Closing remarks and future outlook