Getting started
Introduction
Common Compute is a curated catalog of AI workloads — OCR, text embeddings, background removal, object and face detection, text-to-speech, and HEIC/HEVC conversion today, with more rolling out — each running natively on Apple's Neural Engine or VideoToolbox. Submit a job and get a deterministic quote before anything runs. Tasks execute in native runners on real Apple Silicon Macs, individually owned by network providers.
This guide walks you from install to your first production job in under ten minutes. If you already have an OpenAI, Cohere, or AWS Transcribe integration, you can port in a single line change.
Common Compute is in open alpha. The API is OpenAI-compatible at the wire level for the live workloads; anything marked 'Soon' in the catalog fails fast with a no_capacity error (and an automatic refund) instead of returning placeholder output.
What this is good for
- High-volume batch inference where seconds of latency is acceptable
- Workloads that dominate your AI bill today
- Pipelines with deterministic compute needs and tolerances for retries
What this is not good for
- Sub-100ms realtime inference (use your existing provider)
- Training large models (we dispatch inference only)
- Stateful, long-running processes (tasks are bounded)