SmolVLM2 2.2B on live video streams
SmolVLM2 2.2B · Hugging Face · Tiny VLM · Open-weight
SmolVLM2 2.2B is Hugging Face's smallest video-capable vision-language model, built to run comfortably on a laptop CPU or a single consumer GPU. At 2.2B parameters it trades some accuracy for extreme efficiency, making it a reference point for how much capability fits in an edge-sized model. SmolVLM2 2.2B is not currently in Overshoot's live model catalog; when a tiny model like this is available through the API, it sits at the fast, low-cost end of the lineup: publish a camera feed over WebRTC, reference a frame with an ovs:// URL, and get a lightweight answer back quickly.
SmolVLM2 2.2B is not currently in the live Overshoot model catalog (verified against the live model catalog on 2026-07-14). Availability changes over time - query GET /v1beta/models for the current list.
- Developer
- Hugging Face
- Parameters
- 2.2B
- Context window
- 16K tokens
- License
- Apache 2.0
- Released
- Feb 2025
- Inputs
- Text, images, short video clips
- Overshoot availability
- Not in live catalogas of 2026-07-14
What SmolVLM2 2.2B is good at
SmolVLM2 2.2B is built for places where a full-size vision-language model does not fit: battery-powered devices, kiosks, and browser-only deployments. It handles captioning, short video question answering, and basic scene description while keeping memory and compute demands low enough to run on CPU-only hardware.
Because it was trained with video clips rather than only single images, SmolVLM2 tracks simple temporal changes, like whether an object entered or left a frame, rather than only describing a static snapshot. That makes it a reasonable default for lightweight live-video triage before escalating to a larger model.
- Real-time captioning and scene description on modest hardware
- Short video clip question answering with basic temporal awareness
- A low-cost first pass before routing to a larger model
SmolVLM2 2.2B and the Overshoot workflow
SmolVLM2 2.2B is not currently in Overshoot's live model catalog; GET /v1beta/models returns the current list. The workflow it would slot into is the standard one: Overshoot publishes a camera or screen share over WebRTC through LiveKit, then a chat-completions request points at that stream with an ovs:// reference, either the latest frame or a short recent segment, with no separate upload step.
Responses from live models stream back over SSE, and thread_id keeps a prompt cache warm across repeated questions against the same stream. A model as small as SmolVLM2 is the kind that suits high-volume polling loops, like a monitoring dashboard asking the same question every few seconds, where a larger model's cost would add up quickly.
SmolVLM2 versus larger edge models
SmolVLM2 2.2B is the efficiency floor among small video-capable models. Gemma 4 26B-A4B and Qwen2.5-VL 7B answer more complex, multi-step questions about a scene, but cost more per request and carry more latency. SmolVLM2 is the right starting point when the task is simple and volume is high, with an easy upgrade path if answers need more nuance.
Frequently asked questions
Can SmolVLM2 2.2B analyze live video?
SmolVLM2 2.2B was trained on video clips and handles frames sampled from a live feed, but it is not currently in Overshoot's live model catalog. Live models answer questions about a WebRTC stream through ovs:// URLs, streaming answers back over SSE. The catalog changes over time, so check GET /v1beta/models.
Is SmolVLM2 2.2B open source?
Yes. SmolVLM2 ships under the Apache 2.0 license from Hugging Face, so weights are freely downloadable and usable commercially, including on CPU-only or single-GPU hardware.
Is SmolVLM2 2.2B available on Overshoot?
Not currently. SmolVLM2 2.2B is not in Overshoot's live model catalog today. The catalog changes over time, so check GET /v1beta/models for the current list of hosted and passthrough models.
What is SmolVLM2 2.2B best used for?
SmolVLM2 works well for lightweight, high-frequency tasks: captioning a feed, flagging simple scene changes, or answering short factual questions about what a camera currently sees. For multi-step reasoning or detailed document analysis, a larger model is a better fit.