Google Antigravity SDK Adds Local Offline Gemma 4 Support via LiteRT
Google has updated the Antigravity SDK to support local offline execution via Google AI Edge's LiteRT engine, running Gemma 4 26B A4B with zero API costs.
On September 23, 2026, Google Group Product Manager Sachin Kotwani and Principal Engineer Taylor Mullen announced on the official Google Developers Blog that the Antigravity SDK now natively supports local AI model execution workflows. Powered by Google AI Edge's high-performance on-device LiteRT (LiteRT-LM) inference engine, the update allows developers to run open-weight models such as Gemma 4 26B A4B directly on local workstations with zero API token costs, complete data privacy, and full offline resilience.
Image source: Google Developers Blog / Google AI Edge
On-Device LiteRT Runtime and Dual Agent Configuration Modes
The updated Antigravity SDK provides two primary configuration interfaces to accommodate different developer setups. The first is LiteRTAgentConfig, which directly integrates Google's lightweight on-device runtime, while the second is LocalOpenAIAgentConfig, intended for external local server compatibility.
- LiteRTAgentConfig: Executes pre-packaged
.litertlmmodel checkpoints directly on the host machine. The runtime automatically spawns and manages a local loopback server, detecting hardware accelerators including Apple Silicon Metal, NVIDIA CUDA, NPUs, or fallback CPUs to select the optimal compute path. In addition, LiteRT applies a.lightweight()preset tuned for resource-constrained edge environments. This preset restricts the built-in toolset to core coding operations, streamlines system instructions for smaller context windows, and configures context compaction thresholds specifically calibrated for LiteRT's 64k token KV-cache. - LocalOpenAIAgentConfig: Serves as a versatile connector for developers already running OpenAI-compatible local endpoints, such as Ollama, LM Studio, or vLLM instances.
Gemma 4 26B A4B Checkpoint Deployment and CLI Workflow
The flagship local model highlighted by the Antigravity team is the Gemma 4 26B A4B instruction-tuned model from Google DeepMind. Developers can pull model weights and establish a ready-to-run local checkpoint directly from Hugging Face using Google AI Edge's litert-lm CLI utility:
litert-lm import \
--from-huggingface-repo=litert-community/gemma-4-26B-A4B-it-litert-lm \
gemma-4-26B-A4B-it-gpu.litertlm \
gemma4-26b
This import command downloads an optimized checkpoint of approximately 16.8 GB and registers it locally at ~/.litert-lm/models/gemma4-26b/model.litertlm. Once imported, developers specify the path using the SDK's model_path parameter to immediately initialize local agent sessions. Developers can also use the standalone litert-lm run command to test inference capabilities, speculative decoding, and multimodal reasoning across Linux, macOS, and Windows (WSL) environments.
Offline Privacy, Hardware Requirements, and Practical Boundaries
Running agentic workflows entirely on local hardware delivers substantial privacy and operational advantages. Prompts, indexed codebase context, and tool-call arguments never traverse third-party cloud infrastructure. This makes the local Antigravity pipeline suitable for air-gapped development, strictly regulated industries, and enterprise environments with sensitive intellectual property constraints.
However, the documentation also outlines clear operational boundaries. Comfortably hosting the Gemma 4 26B A4B checkpoint requires dedicated hardware, with at least 24 GB of GPU VRAM or unified system memory recommended. Furthermore, compared to massive cloud frontier models, local edge execution operates within tighter memory and context bounds (such as the 64k LiteRT KV-cache). For complex engineering tasks, Google recommends a hybrid workflow where daily coding assistance and routine tool interactions run locally at zero cost, reserving cloud endpoints for wide-scope architectural refactoring.