Moondream 2 on real-time video

What Moondream 2 is good at

Moondream 2 packs captioning, general visual question answering, object detection, pointing, and gaze detection into fewer than 2 billion parameters, which is unusually broad coverage for a model this small. It has become a default choice for hobbyist and embedded projects that need real vision capability without a GPU-heavy deployment.

Because it supports pointing and gaze detection alongside standard VQA, Moondream handles tasks a captioning-only model cannot, like locating an object in a frame by coordinates or estimating where a subject is looking. Its small footprint keeps inference cheap even at high request volume.

• Object detection and pointing to specific coordinates in a frame
• Gaze detection alongside standard captioning and VQA
• Embedded and hobby robotics deployments with tight compute budgets

Moondream 2 and Overshoot's real-time API

Moondream 2 is not currently in Overshoot's live model catalog, which you can confirm at any time with GET /v1beta/models. The workflow it would slot into is the same one every Overshoot model uses: publish a camera feed over WebRTC through LiveKit, then send a chat-completions request that references the stream with an ovs:// URL, either the latest frame or a bounded recent segment from the 600-second retention window.

If a model of Moondream's size were serving live-stream requests, its small footprint would keep per-request cost low even when polling frequently, and a reused thread_id would keep a prompt cache warm across repeated questions against the same stream, which matters for robotics-style loops that ask the same pointing or detection question on every frame.

Moondream 2 versus other tiny VLMs

Moondream 2 and SmolVLM2 both target the sub-3B tier, but Moondream leans further into structured outputs like pointing and object detection, while SmolVLM2 leans toward video-native captioning. Teams building a robotics or coordinate-based application typically reach for Moondream first; those doing lightweight video summarization lean toward a captioning-focused model instead.

Frequently asked questions

Can Moondream 2 analyze live video?

Moondream 2 accepts image inputs, so it can analyze video by reading sampled frames when self-hosted or served by another provider. It is not currently in Overshoot's live model catalog, so it cannot be pointed at an Overshoot WebRTC stream today. The catalog changes over time; check GET /v1beta/models for the current list.

Is Moondream 2 open source?

Yes. Moondream 2 is released under the Apache 2.0 license by M87 Labs, with ongoing releases since March 2024, so the weights are free to download, modify, and self-host.

What can Moondream 2 do besides captioning?

Moondream 2 supports object detection, pointing to coordinates within a frame, and gaze detection, in addition to standard visual question answering and captioning. That combination is unusual for a model under 2B parameters.

What is Moondream 2 best used for?

Moondream fits hobby robotics, embedded vision, and other compute-constrained deployments that need real detection and pointing capability, not just captioning. It trades some raw reasoning depth for a footprint small enough to run on modest hardware.

Related models

• Florence-2
• PaliGemma 2 28B
• Phi-3.5 Vision
• Qwen2.5-VL 7B