Google Adds Local AI Model Support to the Antigravity SDK
Google has updated the Antigravity SDK to support local model execution using LiteRT, enabling offline agentic workflows with Gemma 4 26B.
The Antigravity SDK running a local CLI resource monitor using Gemma 4.
- The Antigravity SDK now supports local AI workflows using Google AI Edge’s LiteRT.
- Users can deploy the Gemma 4 26B model offline for cost-efficient and private agentic tasks.
- The update enables hybrid orchestration, combining cloud-based planning with local execution for privacy-sensitive code tasks.
Local Model Support in Antigravity SDK
Google has announced a significant update to the Antigravity SDK, introducing native support for local AI model execution. By leveraging Google AI Edge’s LiteRT, developers can now run open-weights models, specifically the Gemma 4 26B iteration, directly on their local machines. This development addresses long-standing developer requirements for offline capability, data privacy, and the elimination of API costs associated with cloud-only inference.
Technical Implementation and Requirements
The integration is built upon the LiteRT framework, which allows for efficient model execution on local GPUs. According to the official developer blog post, the setup requires a machine equipped with at least 24GB of VRAM or unified memory to handle the Gemma 4 26B model effectively. The process involves creating a Python virtual environment and installing the google-antigravity and litert-lm packages. Once installed, developers can import the Gemma 4 26B A4B variant using the LiteRT command-line tool.
Once the model is prepared, developers can instantiate a LiteRTAgentConfig within their Python code to direct the Antigravity agent to the local file path. This approach allows developers to retain control over their execution environment, ensuring that code and data remain on their local machine throughout the lifecycle of the task.
Advantages of Local Execution
Running agents locally provides three distinct benefits for enterprise and individual developers:
- Cost Efficiency: By shifting workloads to local hardware, developers avoid API costs and rate limits imposed by cloud infrastructure providers.
- Data Privacy: For developers operating in compliance-sensitive or corporate environments, local execution ensures that proprietary source code or private data never leaves the local machine.
- Offline Resiliency: Agentic workflows can remain functional in environments where internet connectivity is intermittent or unavailable, ensuring consistent performance for terminal-based utilities or automated scripts.
Hybrid Orchestration Strategies
Beyond strictly local operation, the updated SDK facilitates a hybrid architecture, often referred to as the "Architect-Builder" pattern. In this configuration, a more capable cloud model—such as Gemini 3.8 Flash—acts as a planner, decomposing complex tasks into smaller, actionable components. These tasks are then distributed to a swarm of local Gemma 4 instances that perform the execution, such as auditing code or generating patches.
This hybrid approach allows for the benefits of large cloud models in strategic planning while utilizing local hardware for token-intensive tasks. In testing scenarios involving vulnerability auditing, this method successfully kept approximately 97% of tokens within the local machine, demonstrating significant potential for privacy-conscious automation.
Extensibility and Compatibility
The Antigravity SDK is designed to be flexible regarding the underlying inference engine. While optimized for LiteRT, the SDK provides support for OpenAI-compatible servers, including tools like Ollama, LM Studio, and vLLM. Through the LocalOpenAIAgentConfig class, developers can swap backends without needing to refactor their agent orchestration logic, tools, or established workflows. This modularity ensures that the SDK can adapt to evolving local inference technologies as they become available.
Enjoyed this?
Get more posts like this delivered to your inbox.
