Cloudflare Workers AI runs open models on GPUs distributed across Cloudflare’s network and returns results through an API call or a binding inside a Worker. There is no GPU provisioning: usage is metered in Neurons, the platform’s billing unit, and the catalog spans text, image, audio and embedding models.
What it does
- Runs inference close to users, on the same network that already serves the site or app
- Calls models through REST or a binding, with no intermediary server
- Applies LoRA fine-tuning and processes batches on selected models
How it works
- A Cloudflare account enables the Workers AI API and binding in the same dashboard
- The catalog is hosted and operated by Cloudflare, with open models from several vendors
- Usage is metered in Neurons, with a daily allowance and per-unit billing on the paid plan
- Features such as asynchronous batch and adapters sit at different maturity stages
Strengths
- No GPU rental and no provisioning queue
- Usage-based billing, with a daily allowance included in the plans
- Direct integration with Workers, Pages and the rest of the edge platform
Points of attention
- The catalog changes often and has scheduled model deprecations
- It differs from Cloudflare AI Gateway, which only routes to external providers
- Frontier models require an active payment method even beyond the included usage