Naluri: instinctive, multilingual decisions in one pass
Give it a state (text or JSON) and a typed question with its options written out. Naluri reads them together and returns a probability per option. The model runs entirely in your browser: nothing you type is sent anywhere. Model: jerichosiahaya/Naluri-279M-encoder (CC BY-NC 4.0) · code + report
Loading…
Answers appear here.
Known limits (zero-shot): routing to your own custom labels can be hit-and-miss, and yes/no questions about
structured JSON (e.g. "is this transfer risky?") react only weakly to the data. Fine-tuning on your own labeled
examples fixes most of this. Probabilities close to each other mean the model is unsure.
Hardware: each model downloads once (cached afterwards): 279M ~365 MB and ~2 GB of memory, 308M ~420 MB and ~2.5 GB, 561M ~870 MB and ~4.5 GB. Use a laptop or desktop browser (Chrome 114+, Firefox 135+, Safari 18.4+); the 561M model needs a desktop with plenty of free memory. An answer takes about 0.3–1 s (279M, 308M) or 1–3 s (561M).
Hardware: each model downloads once (cached afterwards): 279M ~365 MB and ~2 GB of memory, 308M ~420 MB and ~2.5 GB, 561M ~870 MB and ~4.5 GB. Use a laptop or desktop browser (Chrome 114+, Firefox 135+, Safari 18.4+); the 561M model needs a desktop with plenty of free memory. An answer takes about 0.3–1 s (279M, 308M) or 1–3 s (561M).