Two renderers, one model
A brief from the benchmark OpenUI publishes, answered by the same model in each format, streamed into each renderer at the same pace, side by side. Watch how each one draws while the answer arrives.
The demo
Loading both renderers.
What you're seeing
- The same model, the same brief. Nemotron 3 Ultra answered each of the 46 briefs in OpenUI's benchmark in our format and in OpenUI Lang, written for OpenUI's chat library with its own prompt. These are that run's first answers, as the model wrote them.
- The same pace. Both arrive 32 characters every 30 ms, so the longer answer takes longer, as it would from the model.
- Each renderer the way its apps use it. MorphScreen's SDK gets each new piece; OpenUI's renderer gets the response so far and renders it again. Each runs in a frame of its own, so neither one's styles touch the other. Ours runs under a strict content security policy; OpenUI's frame allows the inline styles and eval its renderer uses, as its apps do.
- Measured here, in your browser. When the first words showed, how much the part you're watching moved while it drew (cumulative layout shift, lower is steadier; only Chrome and Edge measure it), the time spent drawing and laying out the page after every piece, and when the last piece was drawn.
OpenUI's side is @openuidev/react-lang 0.3.1 and @openuidev/react-ui 0.17.0 from npm, MIT licence, by Thesys (the licences of everything in the frames). The benchmark measures the same things over 20 answers: every row.