Moondream

AI Infrastructure · Computer Vision · Inference

Acquisition

Prime Intellect acquires vision AI startup Moondream for inference speed

A team of fewer than ten engineers built compiler tricks that make AI models answer in milliseconds, exactly what crowded AI clouds now pay to own.

• Prime Intellect, a San Francisco-based AI cloud company valued at $1 billion, acquired Moondream on October 13, 2026, with the purchase price undisclosed.
• Moondream, founded in Seattle by former Amazon Web Services engineers Vik Korrapati and Jay Allen, builds small open-weight vision-language models for reading images like warehouse boxes or livestream ads.
• Moondream's Photon engine compiles each model into a single GPU program with custom kernels, running more than 60 inferences per second on one Nvidia H100 — fast enough for live video.
• The startup had raised only $4.5 million in a 2024 pre-seed round from Felicis, Microsoft's M12 fund and Ascend, and has fewer than 10 employees, per The Information.
• Moondream 3.1 handles 34 requests per second, about 24 times faster than the next most accurate open vision model, and reaches 80% of Meta's Muse Wonder accuracy at roughly 1/1000th the cost.
• Moondream Cloud's inference and Lens services will shut down on November 13, 2026, with customers migrated to Prime Intellect's platform and unused credits refunded.
• Prime Intellect itself raised $130 million in a Series A led by Radical Ventures in July 2026 and had already surpassed $100 million in annualized revenue before the deal.
• The acquisition signals that AI inference providers are now competing on engineering efficiency — squeezing more speed per chip — rather than on raw compute or model size alone.