OpenAI introduces Support for Local AI Models in the Antigravity SDK

Google shipped a notable update to the Antigravity SDK this week, adding native support for running agentic workflows entirely offline using on-device models such as Gemma 4 26B A4B through the LiteRT runtime. The release also bundles drop-in compatibility with OpenAI-compatible inference servers including Ollama and vLLM, giving developers a unified interface regardless of where their models actually run.

This piece breaks down what changed under the hood, how the hybrid orchestration model works in practice, and who stands to gain the most from running token-intensive tasks locally rather than routing everything through the cloud. Developers building privacy-sensitive applications, cost-conscious pipelines, or air-gapped deployments will find the most to act on here.

OpenAI Support for Local AI Models: the announcement

Google introduced local model support in the Antigravity SDK as part of a broader push toward flexible, hybrid AI architectures. The centerpiece of the release is LiteRT integration, which allows Gemma 4 26B A4B to run directly on-device without a network dependency. Alongside that, the SDK now treats OpenAI-compatible inference servers — Ollama and vLLM are named explicitly — as first-class backends, requiring no custom adapter code.

The framing from the Antigravity team positions this as infrastructure work rather than a new product. Developers who already use the SDK's agent abstractions can point those abstractions at a local endpoint or a LiteRT-backed model without rewriting orchestration logic. The announcement did not specify a standalone release version or a public changelog date beyond the blog post itself.

What actually changed with OpenAI Support for Local AI Models

The most substantive change is architectural: the Antigravity SDK now supports a split-execution pattern in which a lightweight cloud model acts as a planner while a local model handles the heavier, more sensitive portions of a workflow. The canonical example in the announcement is code auditing and patching — tasks that produce large token volumes and often involve proprietary source files that developers may not want leaving the machine.

LiteRT is the on-device runtime doing the heavy lifting for Gemma 4 26B A4B inference. For teams already running Ollama or vLLM on internal servers, the OpenAI-compatible endpoint support means those deployments can slot into an Antigravity agent without any protocol translation layer. Both paths share the same SDK orchestration surface, so switching between local and remote execution is a configuration change, not a code change.

Who should care about OpenAI Support for Local AI Models

Security-focused engineering teams are the most obvious audience. Code auditing workflows that touch sensitive repositories, internal tooling, or regulated codebases have historically required either expensive private cloud deployments or a willingness to accept data-egress risk. Running those tasks through Gemma 4 26B A4B via LiteRT eliminates the egress question entirely by keeping inference on the developer's hardware.

Cost sensitivity is a second driver. Token-intensive tasks — lengthy document review, large-codebase patching, multi-file refactors — can accumulate significant cloud inference bills. Offloading that volume to a locally hosted Ollama or vLLM instance converts a per-token variable cost into a fixed infrastructure cost. Teams already operating internal GPU servers will see the most immediate payoff, but even consumer-grade hardware capable of running a quantized Gemma variant becomes relevant here.

How to try OpenAI Support for Local AI Models

Teams using Ollama or vLLM can start by pointing the Antigravity SDK at their existing server's OpenAI-compatible endpoint. Because the SDK treats that endpoint as a standard backend, no additional configuration beyond the base URL and model identifier should be required. The announcement does not specify a minimum SDK version, so checking the Antigravity SDK release notes for the LiteRT integration tag is the right first step before updating dependencies.

For on-device inference with Gemma 4 26B A4B via LiteRT, developers will need to download the model weights in the format LiteRT expects and confirm their hardware meets the memory requirements for a 26-billion-parameter model at whatever quantization level they choose. The hybrid planner-plus-local-executor pattern the SDK enables is worth prototyping before committing to it: measure actual latency and throughput on representative workloads rather than relying on benchmark figures, since device-to-device variance can be significant at this model scale.

What to watch after OpenAI Support for Local AI Models

The practical ceiling here is hardware. Gemma 4 26B A4B is a large model, and LiteRT's ability to run it efficiently will vary considerably across devices. Quantization options, memory bandwidth, and available accelerators will determine whether on-device inference is genuinely faster than a low-latency cloud call for any specific team's setup. Google has not published benchmarks for the LiteRT path on representative developer hardware, which means early adopters will be generating the reference data.

The broader question is how the hybrid orchestration pattern matures. The current framing — cloud model as lightweight planner, local model as token-intensive executor — is a reasonable starting architecture, but the right split between planner and executor will be application-specific. Watch for additional Antigravity SDK releases that formalize routing policies or provide tooling to profile which portions of a workflow are best handled locally versus remotely.