Breaking

Google Adds Local Gemma 4 Execution to Antigravity SDK for Offline Agents

September 23, 2026 at 16:57 EDT

AI Agents
Software Dev & Coding
Foundation Models

Google is adding local model execution to the Antigravity SDK, letting developers run agentic workflows offline with Gemma 4 26B A4B through Google AI Edge LiteRT and connect the same agent code to OpenAI-compatible local servers. The move targets a core tension in agent development: cloud model capability versus token cost, data exposure, and network dependence.

Sept. 23, 2026 · Google

Offline agents move into Antigravity SDK

Local Gemma 4 26B A4B execution brings agent loops onto the developer machine, with a hybrid path that keeps cloud use limited to planning.

97.2%

demo tokens run locally

cloud planner tokens

16.8 GB

Gemma 4 26B import

24 GB+

recommended memory

Hybrid demo token split

One block versus 35 blocks shows the planning handoff: a small cloud step, then local Gemma workers do the heavy loop.

How the local loop works

Plan with Gemini 3.8 Flash

→

Run Gemma 4 26B locally

→

Audit, patch, critique

→

Test on the machine

Why it helps

No API cost for local loops
Code and prompts stay on-device
Work continues without reliable connectivity

What still limits it

High local memory needs
Best support targets one Gemma file
Authenticated endpoint routing remains a gap

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.