On September 23, 2026, Google announced on its official developers blog that the Antigravity SDK now supports local model workflows, launching with Gemma 4 26B A4B through Google AI Edge's LiteRT. Developers can now run agentic assistants in completely offline environments — code and requests never leave the machine.
The requirements are spelled out precisely. Google recommends a machine with more than 24GB of VRAM or unified memory; setup takes two steps — pip install google-antigravity and litert-lm, then download the LiteRT build of Gemma 4 26B A4B from the litert-community organization on Hugging Face — and a few lines of code start a local Agent. In the official sample, that local agent answers "What files are in the current directory?" — the simplest task, and the most telling one.
Why it matters: a cloud brain with a local workforce
Google is not pitching "local replaces cloud" here, but hybrid orchestration. The official demo shows an Architect-Builder pattern: Gemini 3.8 Flash in the cloud handles planning and scheduling, acting as the architect; a swarm of Gemma 4 26B instances on the local machine does the actual work, acting as the builders. In the demo, auditing and patching three vulnerable modules — auth.py, billing.py, and database.py — happens entirely on-device, with data never leaving the local network.
The division of labor is deliberate. Planning and scheduling need the judgment of the strongest model, so they stay in the cloud; the token-hungry execution stage moves on-device, saving API costs and rate limits while satisfying strict data-compliance requirements. The four reasons Google lists — cost, privacy, offline resiliency, hybrid orchestration — all answer the same question: which stages are worth paying a cloud model for, and which can be solved locally. This is the same trend as transformers' native GGUF support (the bar for running quantized models locally keeps getting lower (/article/2045-transformers-yuan-sheng-zhi-chi-gguf-ben-di-pao-liang-hua-mo-xing-bu-yong-zai-zh)): on-device compute is evolving from "can run a model" to "can run an agent."
The boundaries
The 24GB VRAM requirement rules out most thin-and-light laptops — the unavoidable hardware math of running a 26B-class model locally. LiteRT is the preferred execution backend for now; Google only says coverage of other local models and execution options will expand, with no timeline. And offline does not mean maintenance-free: model files, SDK versions, and the local environment still need looking after.
