Cloudflare Workers AI.
Edge inference and agents on 300-plus points of presence worldwide.
What it is.
Cloudflare Workers AI runs open-source models on the company's network of 330-plus cities. Workers itself is the serverless runtime — V8 isolates with sub-millisecond cold starts. The AI layer added GPU inference in 2023 and broadened in 2024 with Llama 3.3, DeepSeek, and Stable Diffusion. Vectorize handles the vector store; Durable Objects handle stateful agent workflows.
Where it fits.
Applications that need inference at the edge — chat widgets, search ranking, personalization, content moderation. Cloudflare's pricing wins against AWS Bedrock or OpenAI for high-volume, low-cost-per-call workloads. The platform is the default choice when latency to the user matters more than frontier-model quality.
- 330-plus global points of presence — lowest latency in the category
- Pay-per-neuron pricing — cheaper than GPT-4 for equivalent workloads
- Integrated Vectorize and Durable Objects for stateful agents
- Open-source models only — no GPT-4 or Claude on the platform
- V8 isolate runtime limits some Python and Node.js libraries
Frequently asked.
Workers AI or OpenAI API?
Workers AI for high-volume, latency-sensitive, cost-sensitive workloads. OpenAI for frontier-model quality where cost-per-call matters less than capability.
What models does Workers AI support?
Llama 3.3, Mistral, DeepSeek-R1, Stable Diffusion XL, Whisper, and a long tail of open-source models. New models ship monthly.
Can I run my own model on Workers AI?
Yes via the Bring Your Own LoRA program for fine-tunes. Full custom models are not currently supported — use Replicate or Modal for that pattern.
