01
What changed in Antigravity SDK
Google added local-model execution to the Antigravity Python SDK, allowing model inference to stay on-device while the SDK continues to manage agents, tools, policies, hooks, workspaces, MCP integration and subagents.
There are two distinct local routes: an SDK-managed LiteRT-LM path optimized around Gemma 4 26B A4B, and LocalOpenAIAgentConfig for separately managed OpenAI-compatible servers such as Ollama, LM Studio and vLLM.
02
LiteRT + Gemma 4 is the documented optimized path
Google's repository says the LiteRT workflow currently works best with the documented Gemma 4 26B A4B GPU LiteRT checkpoint and cautions that other .litertlm files may not work well or may fail.
The current repository recommends at least 24 GB of VRAM or shared/unified memory and recommends a 64K context size for this path. That is Google's guidance for the documented workflow, not a universal minimum for every local model or backend.
03
Ollama, LM Studio and vLLM use an external server
LocalOpenAIAgentConfig connects Antigravity to an OpenAI-compatible inference server that you start and manage separately. Google's repository names Ollama and LM Studio, while the launch post also names vLLM.
Hardware requirements then belong primarily to the selected model and runtime. Antigravity preserves the agent orchestration layer, but different local models can vary substantially in tool use, context, multimodal support and instruction-following.
04
Offline and privacy boundaries
The documented LiteRT path can run model inference without a cloud model API key once the model and dependencies are available locally. That can reduce per-request API exposure and keep code or prompts on the machine.
An agent is only fully offline when its tools are also local. Remote APIs, web tools or remote MCP servers can still transmit data, so local inference should not be marketed as a blanket guarantee that the entire workflow has no network activity.
05
Hybrid architecture and practical limits
Google demonstrates a hybrid Architect-Builder pattern in which a cloud model plans while local Gemma instances perform much of the execution. Google reports 97.2% of tokens ran locally in that recorded example; this is a vendor demo result, not a general benchmark.
Local execution avoids model API charges for local inference but still consumes hardware, electricity and any external service costs. Choose local, cloud or hybrid architecture based on privacy, model quality, latency, hardware capacity and workload economics.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.