Descriptions:
Gus Martins and Ian Ballantyne from Google DeepMind presented the newly released Gemma 4 model family at the AI Engineer conference, detailing the architecture choices, sizing strategy, and use-case positioning that distinguish Google’s open-weight lineup from its proprietary Gemini models.
Gemma 4 ships in four sizes. The E2B and E4B models target mobile and edge devices — the “E” stands for “effective,” meaning a 5B-parameter model that uses only 2B of GPU memory by storing token-mapping parameters in separate memory. Both smaller models support text, vision, and audio input with thinking, coding, and function-calling capabilities and are designed to run on a Pixel phone today. The 26B mixture-of-experts model provides the compute footprint of a 4B model with significantly greater capability, while the 31B dense model currently ranks fourth and seventh on LM Arena’s open-model leaderboard despite being two to twenty times smaller than competing top-ranked models.
Ian Ballantyne focused on the economic argument for open-model ownership in agentic workloads, citing OpenRouter’s state-of-AI data showing that programming tasks carry among the highest combined input/output token costs of any use case. As agents handle more coding and analysis work, he argued, the cost and data-sovereignty benefits of self-hosted models compound significantly. The talk included live demos of Gemma 4 performing modular code generation and document analysis on local hardware, positioning the models as practical drop-in endpoints for teams that cannot or will not route sensitive data through external APIs.
📺 Source: AI Engineer · Published June 10, 2026
🏷️ Format: Keynote Launch
