Gemma 4: Google's Open Models for On-Device, Offline AI

Melvin Vivas · X video post · 2026-06-25 · 4:07 · 215 views · Open on X

Topics: Industry Trends & Job Market, LLM Fundamentals, Fine-tuning & Model Customization · Level: beginner

Summary

Melvin Vivas shares Google's promo video for Gemma 4 and asks whether Gemma might come to Apple devices. In the video, Google presents Gemma as a family of open models that can run on a phone without an internet connection, with sizes chosen for different devices and tasks. Developers in Africa, Thailand and Peru describe fine-tuning Gemma for local languages and for offline apps in places like schools, hospitals and maternal health care.

Key points

• Most frontier models need an internet connection, which shuts out billions of people. Gemma's goal is strong AI that can run anywhere, on any device.
• Gemma is an open model family. Small sizes are built for mobile, and larger sizes focus on coding and agentic tasks.
• Gemma 4 builds on research done for Gemini 3. It is multimodal and can act as an agent directly on a phone, for example coding basic Android apps.
• For small models, the main goal ('North Star') is intelligence per byte of memory.
• Because the weights are open, you can download the model and fine-tune it, for example for a low-resource language like Quechua or for Thai.
• On-device models allow offline apps that use no mobile data and make no API calls. That suits schools, hospitals and remote areas such as the Andes.
• Example from the video: an offline app used every day to help lower maternal mortality.
• Google says open-source AI is key to making the technology available to everyone and to preserving cultures.

Resources mentioned

• Gemma 4 · tool · ai.google.dev · free
  Google's family of open-weight models in several sizes, built to run on devices and offline, with multimodal and agentic abilities, and open to fine-tuning.
  Also in: Gemma 4 Runs Locally On-Device in the Antigravity SDK (Melvin Vivas on X · notes), On-Device AI: Running Gemma 4 E2B Offline on an iPhone with LiteRT (Melvin Vivas on X · notes), Running Gemma 4 Models Offline on an iPhone (Melvin Vivas on X · notes), Fine-tuning Gemma4-E2B on your own tweet style with Unsloth (Melvin Vivas on X · notes) and 25 more
• Gemini 3 · tool · deepmind.google · check price
  Google's flagship model family. The video says its research was reused to build Gemma.

Try this

• Download a Gemma model and try running it on your own device.
• Try fine-tuning Gemma for a specific language or field that has few digital resources.
• Fine-tune Gemma for an indigenous language such as Quechua and ship it as an offline mobile app for schools or hospitals.
• Build an offline, data-free health assistant app, like the maternal-health example in the video.
• Build a homework-checking assistant that runs on a phone.