Google's Antigravity SDK can now run agents offline on a local Gemma 4 model

The Antigravity SDK, Google's Python library for building on the same agent capabilities as its Antigravity coding tool, no longer needs a cloud model. Google announced on September 23, 2026, that the SDK now runs agents on local models, starting with Gemma 4 26B A4B. The company says this lets an agent work completely offline, with no API costs and no rate limits.

What changed

Local execution goes through LiteRT, the on-device runtime from Google AI Edge. You install two packages with pip install google-antigravity litert-lm, import the model file gemma-4-26B-A4B-it-gpu.litertlm from Hugging Face, and point LiteRTAgentConfig at it. The agent code itself stays the same as with a cloud model.

Gemma 4 26B A4B is a mixture-of-experts model. According to its model card, it has 25.2 billion parameters in total, of which 3.8 billion are active. That split is why Google chose it: the memory must hold the whole model, but each token only runs through a fraction of it.

What it takes

The hardware bar is set by memory, not by the processor. Google names three conditions in its announcement.

• Google recommends a machine with more than 24 GB of video memory or unified memory.
• Support starts with Gemma 4 26B A4B through LiteRT; Google calls it initial support.
• Google's own sample code warns that local inference could take several minutes.

Hybrid setups and other local servers

Google also shows a split setup. In its recorded example, Gemini 3.8 Flash in the cloud plans the work from file names and task descriptions alone and spends 95 tokens, while several local Gemma 4 instances audit and patch three Python modules. The company says no source code left the machine.

According to Google, 97.2% of all tokens in that run, or 3,322 tokens, were processed locally. Beyond LiteRT, the LocalOpenAIAgentConfig setting connects the SDK to any local server with an OpenAI-compatible API, such as Ollama, LM Studio, or vLLM, without changes to tools or workflows.

What is not known yet

Google has not published measurements of how the local agent compares with cloud models on quality or speed. The token figures above come from a single demonstration run that Google recorded itself, not from an independent test.