Player not loading? Watch on YouTube
This tutorial introduces Google's Gemma 4 model family, tries the 26B version in Google AI Studio, then explains how to run models locally with Ollama. The browser trial requires a Google account. The installation walkthrough covers Windows, Mac and Linux, followed by model selection in the app and command-line downloading and chat.
The speaker recommends E4B as a starting point. He describes E2B and E4B as options for devices with limited resources, citing 5 GB of RAM for E2B. His suggested hardware for the larger models is 16 to 20 GB for the 26B mixture-of-experts model and at least 20 GB or a dedicated GPU for 31B. The demonstration uses a Lenovo Legion T7 with 32 GB of RAM and an RTX 4080. A default download shows 9.6 GB, but the speaker does not clearly identify its model size. These hardware claims are guidance from the presenter, rather than measured minimum requirements.
The local LLM tests include school-related writing prompts, extracting receipt details and generating an HTML page whose button changes the background color. The speaker reports that the page works. A transport-cost problem exposes a limitation: Gemma's answer misses the requirement to leave no empty seats. He says Gemini Pro returns nine buses and nine vans for the same problem. He also describes local use as private and available offline after download; the tutorial does not independently audit those claims.
