Google Brings Local AI Model Support to Antigravity SDK with Gemma 4

Google has expanded the Antigravity SDK to support local model execution, letting developers deploy agentic AI capabilities on their own machines without relying on cloud APIs. The initial implementation supports Gemma 4 26B A4B through LiteRT, a runtime optimized for edge inference.

Why Local Execution Matters#

Running agentic workflows locally unlocks several practical benefits:

• Cost savings: No API call fees or rate limits when processing stays on-device.
• Data security: Code and requests remain entirely on your machine, meeting strict compliance and privacy mandates in regulated industries.
• Offline operation: Agentic tasks execute reliably in disconnected environments or locations with spotty connectivity.
• Hybrid flexibility: Developers can combine lightweight local processing with cloud models when additional capacity is needed, balancing efficiency and capability.

Getting Started with Local Models#

Setting up the SDK requires a machine with at least 24GB of VRAM or unified memory. Developers install the package, download Gemma 4 26B, and configure the agent to use a local model path instead of a cloud endpoint.

The SDK also supports any OpenAI-compatible server—such as Ollama, LM Studio, or vLLM—via the LocalOpenAIAgentConfig option, giving teams flexibility to experiment with different inference backends without rewriting agent logic.

Hybrid Architect-Builder Pattern in Action#

Google demonstrated the approach with an example where a cloud-based model (Gemini 3.8 Flash) acts as a planner, while local Gemma 4 instances execute the heavy lifting. In a security audit task spanning three vulnerable Python modules, the setup achieved notable efficiency: the cloud model spent only 95 tokens on strategy and decomposition—without ever uploading source code—while local models ran 97.2% of the total 3,322 tokens entirely offline, completing security fixes while keeping proprietary code on-device.

Another example shows the agent autonomously building a live-updating CLI resource monitor. Given a single prompt, Gemma 4 26B writes a complete Python application using the psutil and rich libraries, generates dependencies, and validates the output—all running locally.