Gus Martins is a Google DeepMind product manager working on the Gemma open-model family, where he focuses on making capable AI systems practical to run, customize and control outside proprietary cloud services. His work spans model efficiency, multilingual applications and the legal terms that determine whether institutions can deploy AI independently.
Martins moved into product management through developer advocacy. In 2024, he worked in Google AI Developer Relations and co-hosted People of AI with Ashley Oldacre. He subsequently held a product-management rotation and was product manager for Gemma models by Google I/O 2025.
He helped develop EmbeddingGemma, a lightweight multilingual text-embedding model for retrieval and on-device applications, and highlighted community adaptations including GAIA, a Gemma model fine-tuned for Brazilian Portuguese. In June 2026, he co-authored the announcement of Gemma 4 12B, a multimodal model designed to run on laptops with 16GB of memory and released under Apache 2.0.
- Digital sovereignty through open models: Organizations should be able to keep sensitive data on their own infrastructure, adapt models to specific needs and maintain access without depending on a hosted provider.
- Intelligence per parameter: Model quality matters in relation to available memory and compute. Effective-parameter configurations can fit phones, while mixture-of-experts architectures make more capable systems accessible on modest hardware.
- Licensing as product infrastructure: A familiar Apache 2.0 license lowers legal and procurement barriers that can prevent institutions from adopting models distributed under bespoke terms.
- Practical multilingual adaptation: Communities can tailor models to their languages, but fine-tuning brings diminishing returns when the base model already performs strongly in that language.
These priorities shape his approach to efficient on-device AI: useful intelligence should fit the hardware, languages, privacy requirements and institutional constraints of the people deploying it.
