Prime Intellect buys Moondream; its cloud closes Nov. 13

Prime Intellect, a San Francisco company that sells AI training and inference, has acquired Moondream, a small maker of open vision models. The deal was announced on October 13, 2026. For developers, the date that matters is November 13: that is when Moondream Cloud, its hosted service, shuts down. Financial terms were not disclosed.

What Moondream built

Moondream makes vision-language models, AI models that look at an image and answer questions about it. Its pitch has always been that small models can beat large ones at narrow jobs. "Moondream is proof that small, efficient models can beat giant ones at specific tasks," co-founder Jay Allen said in Prime Intellect's announcement.

Prime Intellect says its latest model, Moondream 3.1, leads on object detection, aerial images, charts and crowded scenes. It says the model handles 34 requests per second, about 24 times faster than the next most accurate open model. In its own unpublished test against Muse Wonder, Meta's new frontier model, Prime Intellect says Moondream 3.1 reaches 80% of the accuracy at pointing to objects in photos. It does so at about one thousandth of the cost. These are company figures, not independent tests.

The company is tiny. Moondream had fewer than 10 employees and had raised $4.5 million before the deal, according to Crypto Briefing and AlphaSignal. Co-founders Vik Korrapati and Jay Allen join Prime Intellect to lead its inference work.

Why Prime Intellect wanted it

The main prize is software that makes models run fast, not the vision model itself, Crypto Briefing reports. That software is Photon, Moondream's inference engine. Inference is the work of running a trained model to answer requests, and it is where most of the ongoing cost sits.

Photon compiles each model into a single program for the GPU, the chip that runs AI models. Prime Intellect says that lets one Nvidia H100 run more than 60 Moondream inferences per second. AlphaSignal reports about twice the throughput of vLLM, a popular open-source serving engine, on comparable work. The first Photon release under Prime Intellect supports Moondream, Qwen and Gemma models on the H100.

Prime Intellect raised a $130 million Series A at a $1 billion valuation in July 2026, Crypto Briefing says. By then it reported a yearly revenue pace of more than $100 million.

What happens to Moondream customers

Moondream spelled out the shutdown in its own post. "On November 13, 2026, 30 days from today, the Moondream Cloud (inference and Lens) will shut down," it wrote. Here is what changes and what stays:

Item	What happens
Moondream Cloud, including Lens	Shuts down on November 13, 2026
Unspent purchased credits	Refunded, per Moondream
Credits on Prime's platform	Coming "within the next few days," per Prime Intellect
Moondream 3.1	Available through Prime Inference
Moondream models	Stay open, and the brand continues

The two posts describe credits differently. Moondream promises refunds of unspent purchased credits. Prime Intellect says Moondream Cloud customers will get credits on its own platform. Moondream says it will contact customers directly. Urgent questions go to contact@moondream.ai.

What this means for developers

If you call Moondream Cloud from production code, you have less than a month. Find every place your code hits the Moondream API this week, not in November.

You have two paths. The first is to move to Prime Inference, where Moondream 3.1 is already available and the credits should land. The second is to host the open model yourself, since the weights stay open. Self-hosting removes the vendor risk, but you will need your own GPU capacity and serving setup.

Before you switch, benchmark the new endpoint on your own images. Speed and accuracy figures from a vendor's launch post are a starting point, not a guarantee. Also check how the refund and the Prime credits apply to your account before the cutoff.

The bigger lesson is about small hosted model APIs. A startup with fewer than 10 people can be bought and shut down with 30 days' notice. For any model you depend on, prefer one whose weights you could run yourself if the API disappeared.