Moondream

Moondream is the vision-language model line of M87 Labs, Inc., created by developer Vik Korrapati and grown from a solo open-source project into a funded company. Its niche is the opposite end of the spectrum from this wiki's frontier giants: models small enough to run on consumer hardware and edge devices while handling real visual workloads, a category Moondream effectively defined.¹

History

Moondream appeared in early 2024 as a roughly 1.6-billion-parameter open Transformer VLM assembled from efficient components, and its successor Moondream 2 became one of Hugging Face's most-downloaded vision-language models, adopted for captioning, visual question answering, and detection pipelines where GPT-4o-class APIs would be overkill or too expensive. The project's traction brought seed funding and incorporation as M87 Labs, with the model line, cloud API, and the Photon local-deployment stack as the product surface; the models' four structured vision skills return coordinates and counts rather than prose, which is what makes them practical perception components.¹²

Models and position

The current flagship, Moondream 3 Preview, moves the line to a mixture-of-experts design (9 billion total parameters, 2 billion active) with a 32,000-token context window and four native vision skills: object detection, pointing and counting, visual question answering, and captioning, per the company's documentation.² Within this wiki's coverage Moondream sits with EleutherAI, Microsoft's Phi family, and the small-model tradition as proof that useful capability does not require frontier scale; Apple's on-device foundation models make the same bet at platform scale; its shift from permissive licensing toward the Business Source License with commercial-use gating mirrors the wider open-weight monetization turn documented across the ecosystem, from Stability AI's membership tiers to the split strategies of Mistral AI.¹³

Stated mission and record

M87 Labs describes itself on its about page as "a San Francisco-based AI company focused on teaching computers to see," and presents Moondream as "the world's most efficient Vision Language Model (VLM)."⁴ The company's homepage closes on the line "Every machine will see." and advertises the model line as offering open weights with commercial use.⁵ Moondream publishes no separate charter or safety framework on its site; its stated commitments are that efficiency claim and continued distribution of downloadable weights.

As of July 2026 the record matches the efficiency half of the positioning: three generations of small models have shipped with weights available through Hugging Face, sized for consumer and edge hardware, continuous with the original solo open-source project. The openness half is more contested. After seed funding, licensing moved from the permissive terms of the early releases to the Business Source License 1.1 with a commercial-hosting gate, a shift this wiki tracks as part of the wider open-weight monetization turn alongside Stability AI and Mistral AI. Supporters note that weights remain freely downloadable and locally runnable under the new terms; critics of such moves observe that BSL-style licenses fall outside conventional open-source definitions, so the homepage's open-weights framing carries a narrower meaning than the project's permissively licensed early distribution.¹³

See also

• Moondream (model family)
• Hugging Face
• Large language model

References