I put a 45M-parameter LLM on a phone. It was 100% confident about a number it invented.
The model returned 1230.0 . The input said 560.00 . Confidence: 1.0000 . Not 0.6. Not "uncertain." One point zero, with a written rationale explaining how it got there — a rationale that quoted the correct number and then produced a different one. That's the single most useful thing I learned building react-native-needle , a React Native binding for Cactus Compute's Needle 2 — a 45M-parameter…
The author built a 45M-parameter LLM on a phone with remarkable results. Named react-native-needle, this React Native binding for Cactus Compute's Needle 2 model runs entirely on-device. The model is particularly effective at handling clear, single-intent commands, such as adjusting kitchen lights or setting timers. When given a task outside its capabilities, like capitalizing France, it declines appropriately with a reason rather than creating an erroneous answer.
However, the model has limitations. It struggles with multi-intent input, misinterpreting string fields, and occasionally inventing numbers in crowded strings. The model also reports a perfect confidence score of 1.0 for fabricated numbers, so the confidence score is unreliable. The author learned that documenting memory consumption and thoroughly testing the model's output is crucial for proper deployment.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
