Nebius Realtime Webcam

Point your webcam at something and get a running commentary from a vision model. It is exactly as fun as it sounds, and it takes about thirty seconds to get going.

cp .env.example .env      # paste your key into NEBIUS_API_KEY
npm start                 # http://localhost:8080

No dependencies, no build step, no framework. Node 22.9+ and a browser with a camera.

Grab a key at tokenfactory.nebius.com under API keys.

Shout-out

This is a love letter to ngxson/smolvlm-realtime-webcam, which had the original idea and the good taste to keep it tiny: grab a frame, POST it to an OpenAI-compatible /chat/completions endpoint, show the answer, do it again. Go star it. This version keeps the spirit and swaps the local llama.cpp server for a hosted one.

The model

Everything here runs on deepseek-ai/DeepSeek-V4.1-Flash, pinned server-side. The browser cannot change it, because the browser cannot be trusted with anything. If you want to point it somewhere else, there is exactly one knob: NEBIUS_MODEL in .env.

Thinking is switched off, on purpose

V4.1 Flash is a hybrid reasoner, which is wonderful for hard problems and terrible for a webcam loop. Left alone it spends its entire max_tokens budget narrating its own thought process and returns an empty content. The history fills up with (empty response) and you start questioning your life choices.

So every request carries:

{ reasoning_effort: 'none', chat_template_kwargs: { thinking: false } }

Different model families read different switches, so both go out. If the endpoint rejects them, the server retries once without and remembers. Result: reasoning_tokens: 0, a complete sentence every frame, and roughly half the latency.

The connection is kept warm

Node hangs up on an idle HTTP/2 connection after about four seconds. At a 2s or 5s interval that means nearly every frame re-does TCP and TLS before the model even sees your face. So the server pings a token-free endpoint on a 2s timer while frames are flowing, and opens the connection at startup so the first frame is not the slow one.

Measured at a 5s interval, interleaved, five runs each:

Free speed. The ping costs no tokens.

Things that sound like they should help but do not

Measured, so you do not have to:

• detail: 'low' does nothing. Flat 194 image tokens either way.
• Shrinking frames from 512px to 384px does nothing. Still 194 tokens.
• Lowering max_tokens does nothing for speed.

What is left is roughly 900ms of prefill and round trip that belongs to the endpoint, not to this repo. Prompt wording trims the tail: ask for fewer words, get fewer output tokens, finish sooner.

How it works

browser --frame (JPEG data URL)--> server.js --> /v1/chat/completions
        <------- token stream ----           <--

server.js exists for one reason: your API key stays server-side. The browser never sees it. It also serves the frontend, so there is one process to run and one port to remember.

What is different from the original

• The API key is proxied, not embedded. The original talks to a keyless localhost server, which is fine right up until it is not.
• One request in flight at a time. The original uses setInterval, which quietly stacks requests whenever inference is slower than the tick, and against a hosted endpoint it usually is. Here the loop awaits each response, then waits the interval. "As fast as possible" means back-to-back, not on top of each other.
• Answers stream in token by token, with a blinking caret, because watching text appear is most of the fun. Stats report time to first token next to total latency.
• Frames are downscaled before upload, with resolution and JPEG quality exposed in the UI, since both drive image-token cost.
• A history that collapses repeats into a xN badge instead of fifty identical rows.
• Prompt presets for the usual party tricks: describe, objects as JSON, OCR, people counting.

Cost

Every frame is a billed request and images cost input tokens. A 1-second interval is about 3,600 requests an hour, so this is not a thing to leave running overnight. Start at 512px, quality 0.7, 2s interval, and keep an eye on it.

Configuration

getUserMedia needs a secure context, so serve this over localhost or HTTPS. Point it at a bookshelf, a whiteboard, or your cat, and enjoy.

License

MIT