Google announced that its Antigravity SDK now supports running local models. The change lets the high‑level “architect” agent stay in the cloud while “builder” agents run on a developer’s machine. In a demo that patched three vulnerable Python modules, only about 95 of roughly 3,400 tokens were sent to the cloud, with the rest processed locally. The SDK works with Google’s Gemma 4 26B model via LiteRT and also supports OpenAI‑compatible servers like Ollama and vLLM.
Why it matters
It lets software teams keep source code off the cloud while cutting inference costs.
