Nemotron 3.5 Super VL tested on 300 real bills: reasoning lifts accuracy 91% to 98%
Arindam_1729 · x · 2026-10-26
A developer ran NVIDIA's new open-source vision model Nemotron 3.5 Super VL on 300 bills (150 synthetic + 150 real Indonesian receipts) in the open-source cost splitter Splitter. The model misread Rp 40.000 as 40 on 44 of 150 real receipts, yet every downstream check passed — proof that tools do perfect math on wrong numbers. The copyable setup: the model only parses bills into strict JSON and never does math itself, always calling tools; a checkreceipt validator caught all 30 planted wrong totals with zero false alarms; enabling reasoning lifted accuracy from 91% to 98% at 3x slower speed and half a cent per bill; a single prompt rule cut misreads from 44 to 1. Lesson: test on real data you didn't make.
More from coding & agent
- Firecrawl offers 2.5x credits to a16z Speedrun founders, unveils $75M Series B — devdigest · 2026-10-27
- Chums: open-source Apple Watch app puts Claude Code and Codex on your wrist — brandon_galang · 2026-10-27
- Open-source Clairic Studio turns pasted text into typed data and working code in 5 languages — dfinke · 2026-10-27
- Matt Pocock shares expand/contract trick: split one-way-door PRs to de-risk big changes — mattpocockuk · 2026-10-27
- Qdrant builds Universal I/O to run vector search mmap code on object storage — qdrant_engine · 2026-10-27
- theo publishes tstack skills, Garry Tan urges everyone to 'publish your stack' — ycombinator · 2026-10-26