Prime Intellect, the AI training and inference platform, has acquired Moondream, a Seattle startup that makes small open-weight vision-language models. Prime Intellect announced the deal on its blog on October 13, 2026.
Terms were not disclosed. Moondream co-founders Vik Korrapati and Jay Allen are joining Prime Intellect to lead and scale its inference business and to speed up its work on multimodal models, including vision and computer use, meaning AI that operates software through a screen.
Moondream will remain Prime Intellect's brand for small, efficient multimodal models. Prime Intellect said every Moondream Cloud customer will receive credits on its platform within days, which can be spent on Moondream models or on Prime's wider training and inference services.
What Moondream built
Moondream's core product is a family of vision-language models: models that look at an image and answer questions about it, write captions, find objects or point to them. The company's bet was that a small model trained for a specific job can beat a much larger general model at that job while running far more cheaply.
When it came out of stealth in October 2024, its model had 1.6 billion parameters, small enough to run on phones and other devices at the edge, and had been downloaded more than 2 million times with about 5,100 stars on GitHub. Prime Intellect now says Moondream's open-weight models have millions of downloads every month.
Moondream also built Photon, an inference engine that compiles each model into a single program on the GPU, along with custom GPU kernels and the compiler behind them. According to Moondream's own unpublished measurements, Photon runs more than 60 Moondream inferences per second on a single Nvidia H100, which Prime Intellect said is fast enough for live video, and returns answers in tens of milliseconds.
Prime Intellect said the latest model, Moondream 3.1, leads on object detection, aerial images, charts and crowded scenes such as store shelves, and handles 34 requests per second, about 24 times faster than the next most accurate open model. In an internal evaluation that has not been published, Prime Intellect found it reached 80% of the pointing accuracy of Meta's frontier model Muse Wonder at about one-thousandth of the cost.
“Moondream is proof that small, efficient models can beat giant ones at specific tasks. Prime is building for that same future: many specialized models that companies own.”
The founders and the funding
Both founders came from Amazon Web Services. Jay Allen, the CEO, spent seven years at AWS, where he worked on AWS IQ, and before that was CTO at Porch, a senior director of development at Zynga and spent a decade at Microsoft. Vik Korrapati, the CTO, spent nine years at AWS.
In October 2024 Moondream announced a $4.5 million pre-seed round from Felicis Ventures, Microsoft's M12 GitHub Fund and Seattle-based Ascend. At the time it launched Moondream Cloud, a hosted version of its models, and said it planned to hire full-stack engineers in Seattle. No later funding has been reported.
Why Prime Intellect is buying
Prime Intellect sells a stack for companies that want to train and run their own models, covering compute, reinforcement learning, evaluation, fine-tuning and inference. In July 2026 it raised a $130 million Series A led by Radical Ventures, with NVIDIA Ventures, Intel Capital and Dell Technologies Capital, taking total funding past $150 million. It said at the time that more than 6,000 customers used its stack and that demand had passed $100 million in annualised revenue in under a year.
Moondream was one of Prime Intellect's earliest customers. Prime Intellect said both companies believe a model and the infrastructure that serves it have to be designed together, and that inference decides how fast a model responds and what it costs to run. Photon and Moondream's kernel work are being folded into Prime's stack as part of a push for faster serving.
Allen said the Moondream team had been Prime customers long before the deal, and that Photon can now power far more than Moondream's own models while those models get access to Prime's full training stack.
What ships now
Moondream 3.1 is already available through Prime Inference, running on the Photon engine. Prime Intellect also said two other models, Clef and Clef Flash, are live through Prime Inference and OpenRouter. The companies did not say whether Moondream's open-weight releases will continue on the same schedule or under the same licences.
