Google has launched support for local models in its Antigravity SDK, allowing developers to build and run agentic workflows on their own hardware while keeping the option to use cloud-based models for higher-level planning. In a demonstration involving the auditing and patching of three vulnerable Python modules, the company reports that only 95 out of roughly 3,400 tokens were consumed in the cloud for planning, while the remaining 97.2% ran locally. The approach gives developers tighter control over data privacy, inference expenses, network reliance, and how computational work is distributed.
One of the standout features is hybrid agent orchestration, which adapts Google's Architect-Builder pattern to a mixed cloud-local setup. Under this pattern, a primary, high-context AI agent—the architect—operates in the cloud and maintains the macro-level system view, including dependencies, architectural constraints, and detailed specifications or implementation plans. Once the architect defines a plan by devising a strategy and dividing the task into smaller units using filenames and task descriptions, several specialized AI agents—the builders—work in parallel in isolated environments locally to carry out discrete, focused tasks mapped out by the specification. The builders run on-device and handle tasks such as reproducing security vulnerabilities, proposing code fixes, reviewing candidate patches, and running regression tests, while additional specialized local agents validate proposed changes against provided test suites. Google notes that the architect doesn't require access to the source code itself, and no source code was uploaded to the cloud in the demonstration.
The initial rollout centers on running Gemma 4 26B A4B through Google's LiteRT technology, but developers aren't restricted to LiteRT alone. The Antigravity SDK also supports OpenAI-compatible local inference servers through LocalOpenAIAgentConfig, including Ollama, LM Studio, and vLLM. Reddit user ambassadortim commented that "companies can have more control over their AI model," while FollowingTop3534 questioned whether the cloud-token share metrics capture the full picture, suggesting benchmarks for patch acceptance, wall time, and energy consumption alongside cloud-token share, and noting that the recommended 24GB or more of VRAM or unified memory represents "a real gate." User Latter_Crazy expressed enthusiasm, stating "it can do things I could not do before with the cloud rules" and noting that cost, speed, and quality are all improving.
The hybrid orchestration model addresses a key tension for developers: keeping sensitive code and data on-premises while still tapping into powerful cloud models for strategic decision-making. By confining code-intensive stages of the workflow to local machines, the SDK minimizes data exposure and network overhead, while the cloud-based architect handles the cognitive heavy lifting of planning and coordination. The Antigravity SDK differs from the Antigravity IDE, which is an IDE powered by the same underlying agent infrastructure—the SDK lets developers build their own AI agents programmatically rather than simply using them for coding tasks. For organizations weighing the trade-offs between cloud convenience and on-device control, the SDK's approach offers a middle path: strategic planning in the cloud, execution at home. The hardware requirements may limit adoption in some settings, but early community feedback suggests the flexibility and privacy gains are drawing developer interest. Enterprises navigating tighter data governance and cost pressures may find the local-first execution model particularly compelling, especially as model sizes and regulatory scrutiny continue to grow.