Google Dev: Local Models & Fine-Tuning · Tuesday, September 22, 2026 · GDG Tulsa, OklahomAI, and Techlahoma during Tulsa Tech Week

Save this page for the workshop demos and everything you need to keep experimenting afterward. Every path below works from your own accounts. You don't need access to any Techlahoma account or repository.

Pick your path

1. Try the four demos

Use a laptop with a current version of Chrome or Edge. Models download on first use, roughly 300–470 MB each; retrieval with an answer can load two models. Start the downloads before the workshop if you can.

Imported files stay on your device. Reloading clears imported text. Model answers can be wrong: compare them with the supplied sources or expected outputs.

1. Context

Give a model better instructions and useful information. Try changing one instruction, then compare the answer with and without your context.

2. Retrieval

Let a model look things up. Try the same query with word matching, related meaning, and both combined. Compare which sources rise to the top.

3. FunctionGemma fine-tuning

Choose an everyday task and train a small output adapter while the base model stays frozen. Compare the before-and-after results on unseen examples, then try your own input.

4. Tiny-model training in Rust

Start with a small model and editable Rust source. Change one training topic, run it again, and compare the generated fragments.

These are different ways to solve a problem. You do not need to progress through all four for every project.

2. Build your own in Google AI Studio

You need a Google account. There's nothing to install and nothing to import.