Meta's Llama 4 arrives on Google Vertex AI as a managed API

Running Meta's flagship open-source models in production used to mean wrestling with infrastructure. You had to hunt down available GPUs, wrangle dependencies, manage autoscaling, and pray your instances didn't crash mid-inference.

Google Cloud just removed that friction entirely. On April 29, the company announced the general availability of Llama 4 as a fully managed Model-as-a-Service (MaaS) endpoint on Vertex AI. Alongside it, Google also made the Llama 3.3 70B managed API generally available.

Basically, you can now run Meta's most powerful models without ever provisioning a single virtual machine.

Drop-in API calls, zero server wrangling

The launch covers Meta's latest heavy hitters. That includes Llama 4 Scout, an efficient Mixture-of-Experts model designed for multimodal workloads, and Llama 4 Maverick, Meta's top-tier model built for complex reasoning and deep visual understanding.

Instead of managing clusters, developers can hit an endpoint on Google's backbone. Google handles GPU provisioning, software dependencies, and patching behind the scenes.

Pricing shifts to a pure pay-as-you-go model. You only pay for the prediction requests you actually make, rather than burning cash on idle GPUs sitting in your cloud console.

The onboarding is surprisingly frictionless. Once you accept the Llama Community License Agreement inside Vertex AI Model Garden, you can call the model directly. Even better, the endpoint supports OpenAI-compatible ChatCompletion interfaces. If you already have code written with the OpenAI Python SDK, you simply swap the base URL to Vertex AI, pass your Google credentials, and you're good to go.

Why it matters

We think this is a smart, pragmatic play from Google Cloud.

For developers and teams, it eliminates the single biggest barrier to adopting Meta's open ecosystem: operational overhead. If you've wanted to test Llama 4 Scout for lightweight multimodal tasks or Maverick for heavy reasoning, you can now do it this afternoon without negotiating GPU quotas with cloud sales reps. The tradeoff is typical for managed services: you trade bare-metal control for simplicity, and you are subject to Google's requests-per-minute quotas.

For the AI landscape, it shows Google's willingness to play aggregator rather than just pushing Gemini. While Microsoft Azure has long championed both OpenAI and Meta models, Google is making sure developers who prefer open weights don't leave GCP.

When frontier open models become this easy to consume via API, the real moat shifts from model ownership to cloud infrastructure reliability and developer developer experience. Google knows that as long as your data and tokens run on Vertex, it wins either way.