Read the market before you finish the headline.
FinSense scores financial text as bullish, bearish, or neutral in 42 microseconds — on your own device, in four languages, with no cloud round-trip and no one watching your queries.
90.8%
FiQA financial sentiment
233 MB
INT8 ONNX · DeBERTa-v3
42 μs
Inference · Snapdragon X2 Elite
4
Languages — EN · ES · ZH · JA
What it is
A sentiment model that lives on your hardware.
DeBERTa-v3, distilled to the edge. Three-class financial sentiment — accurate, private, instant.
Runs locally as a 233 MB INT8 ONNX model — desktop, server, or modern phone. No round-trip, no rate limit, no per-call fee.
Nothing is sent to a third party. The headline you score never leaves your hardware.
A REST API for teams that want it server-side — same model, same numbers, your infrastructure.
English, Spanish, Chinese, and Japanese — financial sentiment trained per-language, not machine-translated.
Measured, not claimed
Three public benchmarks. Real numbers.
88.2%
hardest split · held-out · ties FinBERT (87.8%) at half the size
98.28%
bullish / bearish / neutral · 2,388
90.84%
headlines + microblogs · 961
Part of the stack
One measured-AI portfolio.
FinSense is one of several production systems. Explore the rest.
Put FinSense to work.
On-device or API. Paper-test first, deploy on your own hardware when you’re ready.
Get notified when new capabilities ship