reflex · a System One model in your browser

Ask a picture, or a paragraph, and get odds back.

Give it some state and a few questions with fixed answers. It answers every question in one pass and says how sure it is about each option. The model runs on your own GPU through WebGPU; nothing is uploaded anywhere. This page runs the smallest model reflex supports, so read it as a demonstration of the mechanism rather than of the quality: the served configuration is a 4B model on a real GPU.

1Pick a preset or write your own state and questions.
2Load the model once (Qwen3.5-0.8B, about 650 MB, cached by your browser).
3Run. Each answer is a probability distribution, not a sentence.
temperature1.0
Model not loaded.

An image replaces the {"type":"image"} entry in the state. Everything else is plain text or JSON.

noul asks yes/no · choice picks one of up to 26 · score rates on 2–10 ordered levels. Same JSON as the Python server.

Answers
Nothing yet. Load the model and press Run.
Or: triage a pull request

The same idea applied to code review. The page fetches a public GitHub PR, asks five questions about the PR and six about each diff hunk, and lists the hunks a reviewer should actually read. Numbers only; a reasoning model or a human writes the comments. The full-size version is examples/pr_review.py.

hunks10
Stays in this browser, sent only to api.github.com. Needed when a shared network exhausts GitHub's 60 anonymous calls/hour.
Load the model first.