Google has added local model support to the Antigravity SDK, giving developers a way to run agentic workflows on their own machines while still using cloud models for higher-level planning when needed. The change is designed to reduce token usage in the cloud, improve code privacy, and let teams decide how to split work between local and remote inference.
Google Antigravity SDK moves more work on-device
The Antigravity SDK is built around a hybrid agent setup that follows Google’s Architect-Builder pattern. In that model, a high-context “architect” agent creates the overall strategy, while narrower “builder” agents carry out the individual tasks.
With the new local model support, Google says the architect can remain in the cloud and focus on planning, task decomposition, and coordination using filenames and task descriptions, without needing direct access to the source code. The code-heavy parts of the workflow then stay local, where builders can reproduce security issues, suggest fixes, review patches, and run regression tests.
That split is intended to make agentic development more practical for real projects, especially where privacy or compute costs matter. By keeping the sensitive stages on-device, developers gain more control over where code and test execution happens.
What changes for developers
Google says the hybrid setup gives teams more flexibility in how they manage:
- Data privacy — source code can stay off the cloud during local execution
- Inference costs — fewer tokens need to be processed remotely
- Network dependence — local agents can continue working with less reliance on external services
- Workload distribution — planning and execution can be separated across cloud and local systems
In practice, that means a developer could use a cloud model to define a patching strategy, then rely on local agents to do the repetitive or sensitive work such as testing and patch validation. Google positions this as a way to reduce the amount of data that has to leave the developer’s environment.
Google says most tokens stayed local in a demo
To illustrate the approach, Google described a demonstration in which it audited and patched three vulnerable Python modules. In that example, the company reported that only 95 of roughly 3,400 tokens were consumed in the cloud for planning, while 97.2% of the tokens were processed locally.
Google also said no source code was uploaded to the cloud in that workflow. That claim is likely to matter most for organizations that want AI-assisted coding without sending proprietary code into a hosted model environment.
The announcement frames token efficiency as more than a cost metric. It also serves as a proxy for how much of the AI workflow can be kept under the user’s control, particularly when patching security issues or working with internal codebases.
Gemma 4 26B A4B is the first implementation
The initial implementation centers on Gemma 4 26B A4B running through Google’s LiteRT technology. But Google is not limiting the SDK to that stack. The Antigravity SDK also supports OpenAI-compatible local inference servers through LocalOpenAIAgentConfig.
According to the source material, that means developers can use local setups such as Ollama, LM Studio, and vLLM. The broader message is that the SDK is meant to fit into existing local model workflows rather than forcing teams into a single runtime.
This flexibility could make it easier for teams to experiment with different local models, hardware profiles, or deployment patterns without changing the overall agent orchestration approach.
Why the local option matters now
Interest in agentic coding tools has grown quickly, but so has concern over the tradeoffs. Cloud-based agents can be convenient, yet they also raise questions about cost, latency, privacy, and the amount of context that must be shipped to remote systems.
Google’s local model support addresses those concerns by pushing more work toward the developer’s machine. That may be especially useful for organizations with sensitive code, for teams that want tighter control over AI model behavior, or for users who need to balance cloud spend against local hardware investment.
At the same time, the approach is not entirely cloud-free. Google’s design still uses remote planning for the architect role in the example it described, so the SDK is best understood as a hybrid system rather than a full offline replacement.
Community response focuses on control and benchmarking
Commentary around the announcement reflected both enthusiasm and caution. Reddit user ambassadortim said companies can have more control over their AI model and when to change it, or at least more control over that process.
Another user, FollowingTop3534, argued that cloud-token share alone does not tell the full story. They said any useful benchmark should also include patch acceptance, wall time, and energy use, noting that “97.2% of tokens ran locally” does not by itself show whether the run was faster or cheaper on the user’s hardware. They also pointed to the reported 24GB+ VRAM or unified memory requirement as a practical barrier.
A third commenter, Latter_Crazy, expressed excitement about the added capabilities and said the system can do things they could not do before under cloud rules, adding that cost, speed, and quality are all improving.
Antigravity SDK is not the same as Antigravity IDE
Google also clarified an important naming distinction: the Antigravity SDK is not the same as the Antigravity IDE. The IDE is powered by the SDK and shares the same underlying agent infrastructure, but the two products serve different purposes.
The SDK is intended for developers who want to build their own AI agents programmatically, while the IDE is positioned as a coding environment for using that infrastructure more directly. In other words, the SDK is the foundation layer, and the IDE is one application built on top of it.
For developers evaluating AI coding tools, that distinction matters. The SDK may appeal to teams that want to embed agent behavior into custom workflows, while the IDE is more likely to suit users looking for an out-of-the-box experience.
Source: Original report
Was this helpful?
Explore more: AI Automation Services More Software Development Tech News
Last Modified: October 14, 2026 at 10:33 pm
3 views
