Google announced the winners of its Gemma 4 Good Challenge on August 24, including projects that ran AI models on everyday or resource-constrained hardware. Its winner list described applications for classrooms, wildlife research, and physical assistance.
Three projects made the offline emphasis concrete:
- Trido: A voice-controlled digital whiteboard intended to help educators, including those with physical disabilities, manage classroom instruction.
- PenguinAgent: A research and analysis tool for wildlife ecologists working in remote environments, using local models and field material.
- CodeBuddy: A tutoring project in which students could photograph handwritten Python code for local transcription and debugging.
Google's Gemma 4 model documentation shows how much the model choice can affect that hardware question. Its approximate four-bit inference figures are 2.9GB for E2B and 17.5GB for the 31B model, with actual requirements depending on tools and environment. A demonstration described only as running Gemma leaves out a configuration difference large enough to change which device can be used.
Those estimates concern the model. Audio capture, image processing and the project's own data still need resources. The complete application can therefore require more memory than the model alone.
The projects also illustrate why the location of computation can shape a design. A classroom with unreliable connectivity needs a different failure plan from a permanently connected office. A field researcher may need to examine material before a connection is available to upload it.
Offline operation should be checked for the whole task. A model can run locally while another part of the software still needs a cloud service for sign-in, speech processing, document search, or an update. An offline label on one component cannot answer those questions for the entire application.
