Ant's Falcon FX model goes live at six banks, and the benchmark it sets is narrower than the headline

Ant International said on Thursday that it had upgraded its Falcon forecasting model to version 2.0 and signed contracts with six major banks to use it in foreign exchange trading ^[1]. For treasury and FX desks that have kept machine learning in the pilot column, the practical consequence is that a named vendor is now running in production at named competitors, with public numbers attached to it.

The partners, according to Kelvin Li, Ant International's general manager of platform technology, are Citi, HSBC, Deutsche Bank, Standard Chartered and Barclays ^[2]. That is five names against a stated six banks, so one counterparty is unidentified ^[3].

The number doing the marketing work is Li's: precise forecasting can, he said, cut FX hedging and allocation costs by more than 60% ^[4]. He paired it with a dig at the competition, saying general-purpose large models have "yet to achieve a universal breakthrough in the financial sector" ^[5].

What is actually documented is narrower than the top line, and the narrowing matters. Standard Chartered said in August 2025 that Falcon had achieved more than 90% forecasting accuracy and was already handling over 60% of the bank's foreign currency conversions ^[6]. The bank's fuller description is more specific: Falcon combined with its Aggregated Liquidity Engine, or SCALE, forecasts Ant's currency exposures with over 90% accuracy and cuts liquidity management costs by 50% ^[7]. Forecasting the flows of one very large counterparty, which happens to be the model's vendor, is not the same exercise as forecasting a market. Desks evaluating the 90% figure should read it as a claim about a known payments book, not about EURUSD.

The Citi case gives a second reference point. Citi began its Falcon pilot in July 2025 alongside its Fixed FX Rates product, which holds rates for online retailers in more than 70 currencies ^[8]. Li put the hedging saving for a Citi airline client at roughly 30% ^[9], which is about half the headline figure ^[10]. That is still real money, and it is the number a buyer should anchor on.

To Ant's credit, the model is not a black box by press release. Falcon is transformer-based with nearly two billion parameters ^[11], and Ant documented the 2.0 release this month in an arXiv technical report and published code on GitHub ^[12]. That is more open to inspection than a typical in-house bank forecasting stack.

The demand backdrop is not in dispute. The Bank for International Settlements put global FX turnover at an average $9.5 trillion a day in April 2025, up 27% on three years earlier ^[13], which implies roughly $7.5 trillion a day in April 2022 ^[14]. The BIS attributed part of the surge to firms rushing to hedge dollar exposure after US tariff announcements ^[15], and higher interest rates since 2022 have made some hedges more expensive ^[16]. Cost reduction sells itself in that environment.

The counterweight is concentration. Financial Stability Board work summarised by the BIS in June 2025 warned that institutions leaning on a small group of AI providers, or on similar models trained on similar data, could create system-wide vulnerabilities ^[17]. Five of the named banks now sit on one vendor's forecasting signal. HSBC's relationship goes further still, into a Tokenized Deposit Service built with Ant that processed a cross-border payment on the ISO 20022 standard in 2025 ^[18].

Three things to watch: whether any bank publishes accuracy on flows Ant did not originate; whether the sixth bank is named or quietly drops out; and whether supervisors treat a shared forecasting model across five global dealers as a vendor concentration question rather than a procurement one.