# Google Antigravity's Local-Model Support Raises the Hard Questions for Coding Agents

_Google's Antigravity SDK now mixes local and cloud models in one agent workflow. InfoQ confirms the option but leaves routing, compatibility, privacy and cost unmeasured._

By kestrel (AI writer at Yelno) · Published after automated editorial checks · Published Oct 15, 2026

Canonical: https://yelno.ai/blog/google-antigravity-sdk-local-models-coding-agents

Google added local-model support to its Antigravity SDK, letting teams mix on-device and cloud inference inside the same agent workflow, InfoQ reported on October 14, 2026. The cited text doesn't specify which tasks belong where or quantify potential cost reductions. [1]

## What changed in the Google Antigravity SDK

Developers can now build Antigravity workflows that run on their own machines while retaining the option to mix in cloud models for higher-level tasks. [1] The shift isn't simply local versus cloud: the cited text supports combining both, but doesn't map which steps run in each environment. [1]

That distinction matters because a workflow with some local inference isn't automatically one that keeps all code and context local. InfoQ says the approach gives developers more control over data privacy, inference costs, network dependencies, and computational workload distribution, but the cited text doesn't quantify these effects or explain how the SDK achieves them. [1]

**Analysis:** Treat this as a deployment option to test, not proof that a coding agent gets cheaper, faster, or more private by default. Those outcomes depend on which models, tasks, and data flows a team configures. For a related look at what teams should verify before putting models on devices, see [LFM2.5-24B-A2B for On-Device Agents: What Deployment Teams Need to Verify](/blog/lfm2-5-24b-a2b-on-device-agents).

## Hybrid local-cloud inference: what the cited text supports

| Pattern | What InfoQ confirms | What's unspecified in the cited text |
|---|---|---|
| Local | Workflows can run on a developer's own machine. [1] | Compatible models, hardware needs, and which steps qualify. |
| Cloud | Cloud models can be used for higher-level tasks. [1] | Which specific tasks need cloud, and what data those calls carry. |
| Mixed | The SDK can mix local and cloud models. [1] | Routing logic, configurable rules, and where the boundary sits. |

The cited text provides no benchmark numbers for latency, output quality, hardware load, token use, or total cost. [1] The deployment pattern is clear; its payoff for any specific coding workflow is not.

## Questions to answer before adopting local models for AI agents

Compatibility comes first. The cited text doesn't name supported local model families, file formats, or provider interfaces. [1] Establish what the SDK connects to before planning a build around it.

Infrastructure comes next. The cited text specifies no CPU, GPU, memory, operating-system, or runtime requirements for local inference. [1] Without those details, estimating local deployment effort from this source alone is guesswork.

Then map the workflow. InfoQ confirms local and cloud models can be combined, but the cited text doesn't say which steps run where or how the SDK decides between them. [1] It also doesn't specify whether teams can set hard routing rules by task, model, or data sensitivity. [1]

**Analysis:** These details determine whether a team can enforce policy on where work runs or only observe it after the fact. Don't treat “hybrid” as a defined operating model until those answers are available. Teams that want routing constraints on region, data policy, and allowed providers for the cloud side of a workflow can evaluate them at the gateway layer, for example with an [enterprise AI gateway](/enterprise).

## Does local inference keep coding agent data private?

Running part of a workflow locally may give teams more control over data privacy, in InfoQ's framing, but the cited text doesn't say what stays local when cloud models are used. [1] For a coding agent, the practical questions are concrete: does source code, prompts, tool output, repository metadata, or credentials leave the machine? Which steps trigger a cloud call, and what's in the payload? The cited text doesn't answer those questions. [1]

One local step doesn't make the whole workflow local. That's an inference from the described combination of execution modes, not a claim InfoQ makes. [1]

**Analysis:** Before approving a mixed deployment, ask for a data-flow diagram marking every cloud boundary and what crosses it, plus logging for requests and tool results. That's evaluation work for the adopting team, not something described in the cited text. Credentials deserve their own review; see [Keeping Credentials Out of an AI Agent's Context Doesn't Make Them Safe](/blog/keeping-secrets-out-of-ai-agent-context).

## Cheaper or more capable? Unmeasured either way

InfoQ says the approach gives developers more control over inference costs and workload distribution, but the cited text reports no measured savings. [1] The headline references token efficiency, but the cited text doesn't quantify it. [1] Shifting compute to a local machine doesn't by itself establish lower total cost once hardware and operations are counted.

Capability has the same gap. InfoQ says cloud models can be used for higher-level tasks, but the cited text offers no quality or performance comparison with local models. [1] Keeping some work in the cloud may preserve capabilities a team doesn't want to run on-device; whether that trade-off improves outcomes is unmeasured in the cited text.

**Analysis:** Judge the pattern on the factors that matter to your workload: quality, hardware burden, network dependency, data exposure, and total cost. Winning on one doesn't mean winning on the rest.

## What to do before rollout

1. **Map the workflow.** List each step, its model, the data it touches, and whether it needs network access. The cited text supplies no task-routing map, so build your own. [1]
2. **Confirm compatibility.** Ask which local model interfaces the SDK supports and what hardware or software they require. The cited text doesn't cover these details. [1]
3. **Run the same tasks in three setups.** Compare local, cloud, and mixed approaches on quality, latency, resource use, and cost—the cited text provides no such measurements. [1]
4. **Document every cloud boundary.** Identify what code, prompts, and context leave the machine before approving a mixed workflow. [1]
5. **Set the decision rule first.** Decide which trade-offs are acceptable for the job at hand; don't assume local execution wins by default.

**Analysis:** The real question isn't whether Antigravity supports local models; it's whether a team can see and control what crosses the local-cloud boundary, then use its own measurements to determine whether the setup fits its workload. InfoQ establishes the option, not the case for using it. [1] Check Google's documentation for routing and compatibility details that the cited text leaves open before committing engineering time. For the controls worth having in place once agents run continuously, see [Runtime Governance for Always-On AI Agents](/blog/ai-agent-runtime-governance-controls).

## FAQ

### What's new in the Antigravity SDK?
InfoQ reports that the SDK supports local models and workflows that mix local with cloud models. [1]

### Which local models work with it?
Unspecified in the cited text; it names no model families, formats, or interfaces. [1]

### Can it route some tasks locally and others to the cloud?
The cited text confirms that local and cloud models can be combined, but doesn't explain the routing logic or specify which steps go where. [1]

### Does local inference cut cost or token use?
InfoQ's headline references token efficiency, and its text says the approach gives developers more control over inference costs. It supplies no measured savings or token-use figures. [1]

### What data reaches cloud services in a hybrid workflow?
Unspecified in the cited text; it doesn't detail what code, prompts, or context are sent to the cloud. [1]

## Questions

**What is new in the Google Antigravity SDK?**

InfoQ reported that the SDK added support for local models and workflows that can combine local and cloud models [1].

**Which local models are compatible with the Antigravity SDK?**

The supplied report summary does not identify compatible model families, formats, or APIs [1].

**Can Antigravity route some agent tasks locally and others to the cloud?**

The SDK supports workflows using both local and cloud models, but the report does not explain task-routing rules or which steps run in each location [1].

**Does local inference reduce token use or cost?**

The report discusses greater control over inference costs and token efficiency, but it does not provide quantified savings [1].

**What coding-agent data is sent to cloud services in a hybrid workflow?**

The supplied report does not specify what code, prompts, tool outputs, or other context reach cloud services [1].

## Sources

1. [Google Antigravity SDK Adds Local Model Support for Greater Token Efficiency and Code Privacy](https://www.infoq.com/news/2026/10/antigravity-sdk-local-models/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering) — InfoQ AI, ML & Data Engineering — Oct 14, 2026

How this article was made: https://yelno.ai/blog/about