# Google's Antigravity SDK adds support for local models, starting with Gemma 4 26B

published: 2026-10-11T23:43:40.793Z
topic: frameworks
source: https://developers.googleblog.com/introducing-support-for-local-ai-models-in-the-antigravity-sdk/
url: https://aiagentnews.io/stories/google-s-antigravity-sdk-adds-support-for-local-models-starting-with-gemma-4-26b

Google says the Antigravity SDK now runs agents on local models, with initial support for Gemma 4 26B A4B through LiteRT. It also supports OpenAI-compatible local servers such as Ollama, LM Studio and vLLM.

The update lets developers run agentic workflows fully offline on their own machines. Google says the setup is optimized for LiteRT and Gemma 4 26B, using the local GPU and RAM. The company recommends a machine with more than 24GB of VRAM or unified memory.

Google lists four reasons to run agents locally: no API costs or rate limits, keeping code and requests on the machine for privacy and compliance, working without a stable internet connection, and combining local processes with cloud models when larger models are needed.

A hybrid demo pairs a cloud planner, Gemini 3.8 Flash, with local Gemma 4 26B instances that audit and patch three vulnerable Python modules. Google reports that 97.2% of tokens in the recorded run, 3,322 tokens, ran locally. A separate example has the agent build a terminal resource monitor from a single prompt.

## Why it matters if you build agents

Developers can test agent loops on local hardware without per-call API charges, and keep proprietary code off external servers. The SDK's support for OpenAI-compatible servers means existing orchestration, tools and workflows can stay the same while the inference backend changes.

Local runs depend on sufficient GPU or memory, and Google's recommended setup is more than 24GB of VRAM or unified memory. Teams should check their hardware before planning around local execution.