Google’s Gemma 3 and Gemma 3n large language model (LLM) families offer some of the most powerful LLMs yet for their respective sizes. Moreover, they are multilingual and multimodal by design—Gemma 3 models can handle up to 140 languages, while Gemma 3n supports arbitrary combinations of text, audio, and image inputs by default. In this talk, Tyler Mullen, Staff Software Engineer on the MediaPipe team at Google, discusses how his team leveraged the same technology powering Chrome Built-in AI to bring these model families to the web. He walks through some of the challenges they overcame in order to achieve top speeds while running fully on the user’s device. Then, he covers their straightforward API for running these models with only a few lines of code, allowing anyone to quickly and easily build powerful WebAI applications. He concludes by showcasing several practical (and fun) examples of this tech in action.
Resources:
Docs → https://goo.gle/mediapipe-llm-inference-web
HuggingFace Spaces →https://goo.gle/4hLN39H
See more Web AI talks → https://goo.gle/web-ai
Subscribe to Chrome for Developers → https://goo.gle/ChromeDevs
Event: Web AI Summit 2025
Speaker: Tyler Mullen
Products Mentioned: AI for the web, Gemma 3, Gemma 3n
#ChromeforDevelopers #WebAI