Inference at the edge. Milliseconds from any user.

Edge inference is the right choice for any user-facing feature where latency is perceptible — smart autocomplete, real-time content suggestions, inline sentiment scoring on support ticket input. Workers AI eliminates the round-trip to a cloud data centre.

• Llama 3.1 8B (edge) — Sub-50ms reasoning for common queries at the network edge
• Mistral 7B (edge) — Lightweight general reasoning with fast first-token latency
• Streaming responses — Token-by-token streaming for chat interfaces that feel instant

Whisper Large v3 on Workers AI transcribes audio at the edge — meeting recordings, voice notes, call-centre audio — without audio data leaving the region it was recorded in. Ideal for GDPR-sensitive audio processing pipelines.

• Whisper Large v3 — State-of-the-art transcription, processed in-region
• Multi-language — 99 language transcription with automatic language detection
• Timestamped output — Word-level timestamps for meeting minutes and search indexing

Run a fast lightweight model at the edge before routing to an expensive cloud model. Cloudflare Workers AI acts as a cost gate — simple queries are resolved at the edge, complex ones are escalated to GPT-4o or Gemini 1.5 Pro. This alone can cut inference costs by 40–60%.

• Fast pre-classification — Tag query complexity at the edge, route accordingly
• Intent detection — Identify user intent before invoking a full RAG pipeline
• Cost-gating — Resolve simple queries at edge cost; escalate complex ones

With 300+ global points of presence, Workers AI ensures users in any geography get low-latency inference. Regional data boundaries are respected automatically — EU users' data is processed in EU data centres, US data stays in the US.

• Regional routing — Requests routed to the nearest Cloudflare PoP automatically
• GDPR data residency — EU inference stays within EU; configurable per-region
• Zero egress costs — Cloudflare's flat-rate model eliminates data transfer fees