Ian Ballantyne is a senior developer relations engineer at Google DeepMind who helps developers deploy the Gemma family of open models on phones, laptops, and other infrastructure they control. His work makes advanced AI practical within real constraints: limited memory, operating costs, sensitive data, and the need to inspect what autonomous software actually does.
From Google APIs to open models
Ballantyne was already working on Google’s AdWords API team in 2017, supporting developers building against its advertising platform. His public projects include Google Ads Java tooling, Google Shopping samples, and subsequent experiments with Gemini and generative AI on edge hardware.
He coauthored the Gemma 3n developer guide, contributed to the Gemma 3 technical report, and is among the authors of EmbeddingGemma research. His work on Google’s mobile-actions dataset concerns lightweight models that translate requests into actions on mobile devices. In 2026, he and Omar Sanseviero published a DiffusionGemma developer guide examining parallel text generation, bidirectional context, and deployment across different GPU configurations.
- Open-model ownership has operating costs. Local inference can reduce dependency on hosted services and exploit hardware an organization already owns, particularly for token-intensive coding, translation, and document processing. Ballantyne emphasizes that those benefits must be weighed against electricity, maintenance, uptime, memory, and whether a workload requires immediate responses or can run in the background.
- Hardware fit depends on memory architecture. His Gemma work explains how per-layer embeddings can reside outside scarce accelerator memory, making models more practical on constrained devices. Parameter counts alone cannot establish deployability; quantization, activated parameters, available memory, and the specific task also matter.
- On-device AI can take concrete actions. Through AI Edge Gallery and related work on mobile function calling, Ballantyne demonstrates models selecting skills, opening applications, retrieving local information, and producing small programs directly on a phone.
- Local agent orchestration should remain inspectable. He has demonstrated locally hosted models distributing translation and SVG-generation work among subagents, and coding agents building and debugging browser games. His Antigravity demonstration emphasizes implementation plans, browser inspection, execution traces, and human intervention. For Ballantyne, task-specific model evaluation matters more than generic benchmark standings: the relevant question is whether a particular model completes a particular workflow within its actual infrastructure constraints.
