Prime Intellect acquired Moondream. Founders Vik Korrapati and Jay Allen will lead inference scaling. They will also accelerate Prime’s multimodal model work. Moondream built open-weight vision-language models with millions of downloads. It also built Photon, an inference engine compiling models into single GPU programs. Moondream 3.1 leads on object detection and crowded scenes. It runs at 34 requests per second. It matches 80% of Meta’s Muse Wonder accuracy at 1/1000th the cost. Photon runs over 60 Moondream inferences per second on one H100. Moondream 3.1, Clef and Clef Flash are now live via Prime Inference. Moondream Cloud customers will receive platform credits. Moondream will remain a brand for small efficient multimodal models.
What changed
Moondream operated independently with its own inference engine and model catalog.
What it unlocks
Running Moondream models through Prime Inference with Photon engine integration.
- 34 requests per second for Moondream 3.1
- 24x faster than next most accurate open model
- 80% of Muse Wonder accuracy at 1/1000th the cost
- 60 inferences per second on a single H100
Sources