Constellation brings AI inference to Workers
Cloudflare's developer platform has expanded from compute and storage into databases, streaming, networking, and security. The company has consistently pushed developers toward building full applications on its edge network instead of traditional architectures. Now, with the introduction of Constellation, pre-trained machine learning models and inference tasks can run directly on Cloudflare's infrastructure.
Constellation lets developers execute fast, low-latency inference using pre-trained models natively from Workers scripts. The service supports a broad set of use cases, including image classification, object detection, anomaly detection, text translation and summarization, sentiment analysis, speech recognition, and question answering.
Model management and APIs
Developers can upload any supported ONNX model to Constellation—either training it themselves or fetching a ready-made model from hubs like HuggingFace or ONNX Zoo. Cloudflare will also maintain its own catalog of verified models for developers who prefer not to source and test models on their own.
Wrangler, Cloudflare's CLI, includes Constellation project management through the constellation keyword. The workflow involves creating a project bound to the ONNX runtime, editing wrangler.toml to add the project binding, installing the Constellation client API library, and uploading the chosen model.
Example: image classification
To demonstrate Constellation as a building block of Workers-based apps, Cloudflare released a small image classifier using SqueezeNet, a convolutional neural network pre-trained on ImageNet's million-plus images and capable of classifying images into 1,000 categories. SqueezeNet achieves accuracy comparable to AlexNet while running roughly 3x faster and producing models about 500x smaller, making it suitable for resource-constrained devices.
The Worker reads an image from the request, decodes it (currently PNG-only) into a multidimensional float32 tensor, and passes it to the SqueezeNet model running in Constellation. The returned synset IDs are matched against a class-label mapping file, and human-readable tags are sent back in the response. The demo lists probabilities alongside each classification; for example, a photo might trigger high confidence on "alp" and "convertible" and a lower confidence on "Ibizan hound" when the dog in the image is actually another breed.
The full source for the example is available on Cloudflare's developer documentation.
Transformers and stack interop
Constellation also supports transformer models, the deep-learning architecture commonly used for NLP tasks like translation and summarization. Cloudflare ported Transformers.js, a popular browser-based demo that loads HuggingFace transformer models via ONNX Runtime in WebAssembly, to run on Constellation instead. The port is live at https://transformers-js.pages.dev/.
Because Constellation runs natively on Workers, it can be orchestrated with other Cloudflare products. A Worker could, for example, receive email through Email Routing, run a t5-small sentiment analysis model from HuggingFace in Constellation, and add an X-Sentiment header before forwarding the message to its destination. Email clients like Gmail can then filter messages based on that header—routing angry mail out of the inbox before it arrives.
Availability and roadmap
Constellation enters private beta today. Developers can join the waitlist from the dashboard by opening the Workers tab and requesting access under the Constellation entry; Cloudflare will onboard accounts in batches and notify users by email.
The Constellation developer documentation is already live, and the feature is manageable from both Wrangler and the dashboard UI. Cloudflare plans to increase limits, support more runtimes, and accommodate larger models in response to user feedback.
Notably, Constellation itself was built with the same stack developers use: WebAssembly, Workers, R2, and Cloudflare's own APIs. A follow-up post explaining that architecture is planned. The team will also be based in the #constellation channel on the Developers Discord and on the Cloudflare Community forum.
