Introducing Support for Local AI Models in the Antigravity SDK

The Google Antigravity SDK now lets developers execute offline, agentic workflows locally, using models like Gemma 4 26B A4B via LiteRT. Local models mean no cloud call, no code leaving the machine, and token budgets that don't exist anymore.

Hybrid orchestration: cloud architect meets on-device workforce

The standout pattern is an Architect-Builder split: a cloud model handles planning while a local swarm does the work. In the demo, built with the updated Antigravity SDK, a cloud architect (Gemini 3.8 Flash) acts as planner and conductor while a local swarm of Gemma 4 26B instances handles the heavy lifting entirely on-device.

The demo task: audit and patch three vulnerable modules (auth.py, billing.py, database.py) while maintaining strict data privacy and minimizing cloud token use:

• No code uploaded — Gemini 3.8 Flash plans the strategy and decomposes the work from filenames and task descriptions alone, spending just 95 cloud tokens without any source code ever leaving the machine.
• Autonomous local gauntlet — local Gemma 4 26B models run an adversarial audit loop on the local GPU: reproducing security vulnerabilities, authoring candidate fixes, critiquing patches, and validating against regression test suites.
• Massive cost and privacy wins — in the recorded run, 97.2% of all tokens (3,322 tokens) ran locally and offline without calling a cloud API, producing fully verified, green patches while keeping proprietary code completely secure on-device.

Token-free local utilities

The same SDK also excels at building practical system utilities. Given a single prompt, the agent autonomously writes a Python script using psutil and rich to build a live-updating terminal resource monitor — generating the requirements.txt, and even testing the resulting code to ensure it works — all locally on Gemma 4 26B.

Why this matters

The Architect-Builder split reframes the local-vs-cloud trade: you don't have to choose between "everything private" and "everything smart." The planner (which needs the most intelligence) runs in the cloud with minimal token spend, and the repetitive, code-touching work stays on-device where privacy and cost are best. The 97.2%-local number is the concrete proof that most agent work is execution, not planning — and execution is exactly what local models are now good at.

Bottom line

If you deal with proprietary code or sensitive data and have been avoiding agent tooling because of what it would upload, the Antigravity SDK's local-model support is worth a serious look — the hybrid pattern gives you cloud-scale planning without the code ever leaving the machine.