Prime Intellect Buys Moondream to Own the Fast Path for Running Open Vision Models

Prime Intellect has acquired Moondream, the startup behind a family of small open-weight vision-language models, bringing both the models and the software that runs them under one roof. The deal, announced October 13, closes with the two companies already intertwined: Moondream 3.1 is live on Prime Inference, served by Moondream's own Photon engine.

Co-founders Vik Korrapati and Jay Allen are joining Prime Intellect to lead its inference work. Prime, a San Francisco company that raised an equity round in July at a $1 billion valuation and has since disclosed an annualized revenue run rate above $100 million, was one of Moondream's earliest customers. Terms of the acquisition were not disclosed; Moondream, with fewer than ten employees split between San Francisco and Seattle, had previously raised $4.5 million.

What Prime is really buying is execution technology. Moondream's team specializes in megakernels — a technique that fuses the many independent math operations a GPU performs into one continuous computation, cutting the idle gaps between tasks during inference. Writing megakernels by hand is slow, specialist work; Moondream built a custom compiler that can automatically generate them for many models, along with the GPU kernels that run them.

The stated performance numbers come from Moondream and Prime themselves and have not been independently verified. Prime says Moondream 3.1 handles 34 requests per second on selected visual benchmarks, roughly 24 times the speed of the next most accurate open model. An unpublished internal evaluation, also company-supplied, puts the model at 80 percent of Meta's Muse Wonder accuracy on pointing to objects in photos at about one-thousandth of the cost. Prime adds that Photon serves more than 60 Moondream inferences per second on a single Nvidia H100.

The acquisition lands in a market where neoclouds and inference providers are fighting over who can run open models fastest and cheapest on Nvidia hardware. Most inference vendors hand a large share of their revenue to the cloud operators that own their servers; Prime operates its own hardware, which is one reason its margins differ from peers like Fal, above $800 million in annualized revenue, and Together AI, above $1 billion.

Prime's CEO Vincent Weisser told Bloomberg's Chinese-language service that the team's expertise in optimizing non-text modalities is scarce, and that the company is aiming the technology at customers building robotics AI and computer-use agents — markets other inference providers have not yet served well. Photon is also meant to serve models beyond Moondream's own, which will be the real test of whether the acquisition buys a capability or just a model.

Moondream Cloud customers will receive platform credits over the next few days, and the Moondream name will survive as the brand for its small multimodal models. The larger integration is still underway, with Prime folding Photon and Moondream's custom GPU software into its platform of evaluation, fine-tuning, reinforcement learning and inference services.

The deal is a small one by the standards of this market, but it maps out where the inference business is heading: the money is moving from who owns the weights to who owns the machinery that decides how fast those weights answer and what each token costs.