01
Requirements and the documented model path
Google's current local-model documentation recommends at least 24 GB of VRAM or shared/unified memory and a 64K context size for the documented Gemma 4 26B A4B LiteRT workflow.
The repository says the path currently works best with the specific Gemma 4 26B A4B GPU LiteRT checkpoint. Do not assume every .litertlm file is supported just because it uses the same container format.
02
Install Antigravity SDK and LiteRT-LM
Use an isolated Python environment, install the current google-antigravity and litert-lm packages from Google's documented workflow, then import the supported Gemma 4 LiteRT checkpoint locally.
Package names and SDK interfaces can evolve, so copy current commands from the repository at installation time rather than relying on an old snippet. The model and dependencies must already be present before testing truly offline inference.
03
Use LiteRTAgentConfig for the SDK-managed server
Point LiteRTAgentConfig at the local model path and run it through the Agent lifecycle. The SDK manages the local loopback serving path, so the documented LiteRT configuration does not require a cloud model API key.
Google specifically warns not to point LocalOpenAIAgentConfig at litert-lm serve. Use LiteRTAgentConfig when you want Antigravity to own the LiteRT server lifecycle and use LocalOpenAIAgentConfig for independently managed servers.
04
Test offline behavior and tool boundaries
After the model and dependencies are local, verify inference without cloud model connectivity. If the agent uses remote APIs, web tools or remote MCP servers, those tools can still access the network even though the model itself is local.
Grant workspaces and tool permissions only where the task needs them. A permissive sample policy is not automatically appropriate for a real coding or business environment.
05
Context, performance and common mistakes
The repository documents a 64K LiteRT KV-cache capacity and related context-compaction behavior, plus optional optimizations such as speculative decoding and compilation caching. These controls do not guarantee a particular tokens-per-second rate.
Do not convert Google's memory recommendation into a claim that every 24 GB system behaves identically, and do not describe the workflow as fully private if attached tools transmit data. Benchmark the exact device, model and agent workload.
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.