trionTalk to Trion
DevelopersAnnouncement

Antigravity SDK adds local model support

Google has added local model support to the Antigravity SDK. The release uses Gemma 4 through LiteRT to support agentic work that can run without a hosted model connection.

Trion editorialReported 1 min read

Local execution moves the constraints to the device

The SDK announcement provides a local setup using Gemma 4 26B and recommends a machine with more than 24 GB of VRAM or unified memory for that example. It describes offline operation and hybrid workflows that combine local and cloud processing.

Google added support for local AI models in the Antigravity SDK, using Gemma 4 through LiteRT. The update targets offline and local agent work, with hardware requirements that depend on the chosen model and available GPU memory.
Trion

Local work uses local resources

Example model
Gemma 4 26B
Example hardware
More than 24 GB VRAM or unified memory recommended
Hybrid work
Can combine local and cloud processing
Connected tools
May still contact external services
View data
Local-work componentPublished requirement or limit
Example modelGemma 4 26B
Example hardwareMore than 24 GB VRAM or unified memory recommended
Hybrid workCan combine local and cloud processing
Connected toolsMay still contact external services

Google · published 2026-09-23. Source-bound illustration, not a performance benchmark.

Download image

Hardware needs depend on the model and configuration. Keeping inference local can reduce dependence on API access, but an agent’s tools may still contact external systems when the workflow requires them. Local model execution is one boundary; it does not automatically make every connected action offline.

Original sources

  1. Google: original announcementdevelopers.googleblog.com

Checked 9 Oct 2026 · A manually curated edition. Availability may change; company performance claims are not Trion test results. Editorial method.