Europe 2026
Sovereign Escape Velocity: Ownership w Open Models — Gus Martins, & Ian Ballantyne, Google DeepMind
Overview
This talk introduces Google DeepMind's Gemma family of open models, highlighting their capabilities and the benefits of using open-source AI for greater ownership and customization. The presenters emphasize that while proprietary models like Gemini offer cutting-edge performance, open models like Gemma provide crucial advantages for specific use cases, such as running models on local hardware, handling sensitive data, and adapting models to unique requirements.
Who should watch
- AI Engineers and Builders
- Product Managers
- Developers seeking model ownership and customization
- Those interested in running AI models on edge devices or local hardware
- Teams working with sensitive or proprietary data
Key takeaways
- Google offers two complementary model families: Gemini (proprietary, most intelligent) and Gemma (open, for control and access).
- Gemma models are designed for strong performance with reduced hardware requirements, enabling deployment on a wider range of devices.
- The Gemma family includes models optimized for mobile/IoT (2B, 4B effective parameters) and larger, more powerful versions (26B Mixture of Experts, 31B dense).
- Gemma models demonstrate competitive performance on benchmarks like LMSYS Chatbot Arena, often outperforming larger proprietary models in intelligence-per-size.
- The Gemma license has transitioned to Apache 2.0, simplifying adoption for sovereign institutions and commercial use.
- Open models allow for greater control over data, cost, and model behavior, especially beneficial for agentic tasks with high token generation.
- Models can run on various hardware, from phones and laptops to single GPUs, with considerations for latency, energy costs, and task suitability.
- Fine-tuning Gemma for specific languages or domains, like MedGemma for medical use cases, is feasible and unlocks specialized applications.
Notable quotes
*In some situations, you want to own the model. You want to be able to run on your own hardware. You want to customize it.*
*So we are talking about a disproportionate amount of intelligence per size.*
*The question is, do you need the most intelligent model of the planet to summarize your mail, to do some more minial tasks, to help you code, to do some agentic capabilities that are searching and interacting with docs? Probably not.*
Unofficial community note. Prefer the recording for nuance.