Cassidy Hardin

Cassidy Hardin is a Google DeepMind researcher and technical leader developing Gemma, the family of open models designed to bring advanced AI capabilities to ordinary developers and everyday hardware. She was a core contributor to Gemma 2, a technical lead on Gemma 3, and a contributor to the architectural advances behind Gemma 4.

Hardin studied at the University of California, Berkeley, where she mentored computer science students and received an Outstanding Graduate Student Instructor award in 2021. At Google DeepMind, she helped develop the original Gemma models and RecurrentGemma, which combines linear recurrences with local attention to reduce the memory demands of processing long sequences.

The Gemma 2 technical report credits Hardin among four core contributors; the Gemma 3 report names her among four technical leads. Those generations expanded the models’ attention efficiency, visual understanding, context length, and multilingual capabilities. She also helped introduce Gemma 3 to developers at Google’s Tokyo meetup and supported Google DeepMind’s Genie 2 foundation world-model project.

• Memory-efficient model architecture. Her work combines local-global attention, grouped-query attention, recurrent alternatives, and mixture-of-experts routing to preserve model capability while reducing inference and memory costs.

• Per Layer Embeddings for on-device inference. In her Gemma 4 deep dive, Hardin explains how flash-resident embedding tables reduce pressure on device VRAM, helping smaller models run locally on phones and laptops.

• Adjustable multimodal compute. Gemma 4 lets developers tune image resolution and token budgets for demanding applications such as optical character recognition and spatial understanding; its smaller variants also support speech applications through a conformer-based audio encoder.

• Practical developer access. Apache 2.0 licensing, downloadable model weights, local execution, and deployment through Google AI Studio or Vertex AI give developers multiple paths from experimentation to agentic applications.