Google takes agents offline: the Antigravity SDK now supports local models like Gemma 4 26B

Google has announced that its Antigravity SDK — the tool that lets developers build with the same agentic capabilities that power Google Antigravity — now includes support for local workflows, with a range of models and on-device execution options. Initial support centers on Gemma 4 26B A4B, run through LiteRT, Google AI Edge's runtime, optimized specifically to take advantage of the local machine's GPU and RAM.

With this update, developers can enable fully offline agentic assistance, without relying on a connection to cloud services. Google lists four main advantages of running agents locally: cost efficiency (no API spend or rate limits), privacy (code and prompts never leave the machine, which matters for corporate environments with compliance restrictions), offline resilience (workflows keep running even without a stable internet connection) and the ability to combine local processes, which are more token-efficient, with more powerful cloud models when needed.

To get started, Google recommends a machine with more than 24 GB of VRAM or unified memory. The process involves creating a Python virtual environment, installing the google-antigravity package alongside litert-lm, and downloading the Gemma 4 26B A4B model via the litert-lm import command from the corresponding repository on Hugging Face. From there, a minimal Python script — using the SDK's Agent and LiteRTAgentConfig classes — lets you point to the locally downloaded model, launch a conversation with the agent and receive the response token by token, all without leaving the machine.

One of the most striking examples Google shows is a hybrid orchestration pattern dubbed 'Architect-Builder': a cloud model, Gemini 3.8 Flash, acts as the architect that plans and coordinates strategy, while a swarm of local Gemma 4 26B instances does the heavy lifting directly on the device. In the demonstration, the system is tasked with auditing and patching three modules containing vulnerabilities (auth.py, billing.py and database.py). What matters is that Gemini 3.8 Flash plans the strategy based solely on file names and task descriptions, without a single line of code leaving the machine, consuming just 95 tokens in the cloud. From there, the local Gemma 4 26B models autonomously run an adversarial audit cycle: they reproduce the security vulnerabilities, draft candidate patches, critique those patches and validate them against regression test suites. In the recorded run, 97.2% of all tokens used (3,322 tokens) were processed locally and offline, delivering verified, green patches while proprietary code remained entirely secure on the device.

Google also shows a second, more practical-utility example: from a single prompt, the agent autonomously writes a resource monitor for the terminal, using the psutil and rich libraries to display live CPU and memory usage alongside a table of the five most memory-hungry processes. The agent itself generates the requirements.txt file and tests the resulting code to check that it works, with the entire process run locally on Gemma 4 26B.

Beyond Gemma and LiteRT, the Antigravity SDK also offers direct compatibility with any OpenAI API-compatible server, such as Ollama, LM Studio or vLLM, through the LocalOpenAIAgentConfig setup. This lets developers experiment with different local inference backends without having to modify agent orchestration, tools or workflows they have already built.

Google invites developers to consult the README for the Antigravity SDK for Python and the LiteRT documentation to learn how to run models efficiently at the edge, and encourages them to share feedback and feature requests on the project's GitHub issue tracker.