The Comprehension Race
Two financial headlines, one word apart, mean buy or sell.
A resident 0.8B reads them right, in 84 milliseconds, nothing leaving the box.
▶ Watch a 0.8B Flip Bullish to Bearish on One Word
The same model reads a matched pair of headlines that differ by a single word. The naive baseline collapses them; the resident model flips correctly, and it does it before the market has finished reacting.
resident 0.8B on a single GPU · the label is the read of the text, not the return — comprehension, not P&L · watch on YouTube
"Beats EPS, raises guidance" is a buy. "Beats EPS, cuts guidance" is a sell.
The two sentences are nearly identical, and a naive model collapses them, reading the pair the same and getting one of them wrong. Fine-tuned to read financial events, a 0.8B model tells them apart: across a held-out set of matched twins it goes from flipping one pair in ten to nine in ten, and from a coin flip to 95% on the read.
Comprehension before the market finishes reacting is the edge, and an API that is slow and leaky cannot deliver it: 800 milliseconds of round trip is an eternity, and the request ships your signal to someone else on the way. A resident model reads the event in 84 ms on hardware you control. The model here is distilled from 3,063 verified twins through a collapse-proof pipeline. Honest framing: the label is the read of the text, not the trade return — this measures comprehension, not P&L.
Partner with Voxell
Running retrieval, comprehension, or transactional inference where latency and data sovereignty matter? We work with a small number of partners. Reach out.
Knowledge is precious.