Summary
On July 8, 2026 Cloudflare made Moondream 3.1 available on Workers AI, a mixture-of-experts vision-language model (9B total / 2B active, 32K context) that answers questions about images, captions them, and returns object coordinates and bounding boxes — with first tokens returning in roughly 20-30 ms on Cloudflare's edge GPUs.
What changed
Cloudflare added Moondream 3.1 (@cf/moondream/moondream3.1-9B-A2B) to Workers AI on July 8, 2026 through a partnership with Moondream. The model handles visual Q&A, captioning, pointing (object coordinates), and bounding-box detection, using a mixture-of-experts architecture with 9B total and 2B active parameters and a 32K context window.
Why it matters
Running a compact vision-language model on edge GPUs targets latency-sensitive uses — content moderation, screenshot and UI understanding for agents, and real-time visual checks — where round-tripping to a centralized frontier VLM is too slow or costly. It strengthens Cloudflare's pitch that Workers AI is the place to run small, fast, task-specific models close to users rather than only proxying large hosted models.
Evidence excerpt
Use Moondream 3.1 to query, caption, and return object coordinates or bounding boxes in images — vision requests complete in less than a second. In Cloudflare's testing, first tokens streamed back in roughly 20-30 ms.