85.7% Accuracy and Sub-Second Response: Drex for Finance Launches, and the Comments First Ask How to Backstop the Rest
Nace AI has released Drex for Finance, which it describes as a multimodal model for financial decision-making that can make judgments in under a second. It reports 85.7% accuracy and a 77.3 Finance Index, covering five pipelines: documents, credit, fraud, trading, and insurance and regulation. Some commenters asked about high-frequency trading, while others first asked how the remaining decisions would be backstopped.
Nace AI released Drex for Finance, a multimodal model for financial decision-making. The company says it has been post-trained with the goal of pushing response times in financial scenarios down to sub-second levels. Financial operations require precision and auditability, with zero hallucinations across filings, customer support, credit risk, real-time anti-fraud, quant signals, and compliance. This is a very demanding requirement: errors are not just an experience problem, but also a compliance and capital problem.
The official evaluation covers 48 held-out financial benchmarks and reports 85.7% accuracy on 3,361 decisions, 14.8% higher than GPT-6 Luna. The Finance Index is 77.3, 13.7 higher than Jev 1.14 and 11.5 higher than Drex-Max v1.1. According to the company, the model covers the entire pipeline: extraction, classification, and clustering of financial documents and filings; automated underwriting, PD modeling, and EWS for credit and lending; sub-second transaction scoring, ATO defense, and SAR clustering for fraud and AML; regime detection, meta-labeling, and LOB execution signals for markets and trading; claims fraud detection, sanctions screening, and audit reconciliation for insurance and regtech.
The comparison chart released by the company lists models such as Drex Max Finance v1.1, Drex Max Finance v1, GPT-6 Luna, GLiDE, pplx-decider v1.2, and Jev 1.14. The breakdown includes Finance Index, Accuracy, sentiment and tone, news headlines, banking support intent, documents and extraction, financial QA and documents, credit and risk, monetary policy, ESG and climate, and markets and reasoning. The chart provides score comparisons, but the tweet itself does not detail the full list of 48 benchmarks item by item.
Currently, Drex for Finance is only available in Nace Console, at https://console.nace.ai. The tweet does not provide information on API, pricing, or private deployment; to try it, you currently have to go through Console.
The comments did not focus solely on benchmark scores. Some users argued that multimodal decision models are important for high-frequency trading; some found it interesting to put fraud, credit, and compliance into one model; others noted that beyond speed, the accuracy at that speed is what really matters. Some users directly asked how the remaining decisions would be backstopped given 85.7% accuracy, and which benchmarks were specifically used. The official account did not answer these questions in the tweet.
发布时间: 2026-11-03 09:36