Ant Group's "Two Legs": Ling-3.0-flash-Fin Financial LLM + FalconTST 2.0 Time-Series Model

> TL;DR: On August 28, 2026, Ant Group's Bailing lab officially launched Ling-3.0-flash-Fin, a finance-enhanced LLM based on the Ling-3.0-flash architecture (124B total / 5.1B activated MoE parameters), improved through continued pre-training on financial corpora, domain post-training, and tool-calling optimization. It outperforms similarly sized models on benchmarks such as FinFIRST and scores 41 on the AA Intelligence Index. Model weights will be open-sourced next week, with one month of free API access on OpenRouter. Meanwhile, on August 20, Ant International released FalconTST 2.0 (Falcon Time-Series Transformer 2.0), a time-series forecasting model for cross-border payment FX risk management, achieving SOTA benchmark results of MASE 0.666 and >93% accuracy. It is already deployed by six global banks including Barclays (BARX NetFX), Citi (Fixed FX Rates), Deutsche Bank, and Standard Chartered (SCALE FX). Two legs — text understanding via financial LLMs, and numerical prediction via time-series models — are now both running as of August 2026.

Why Now: Ant Group's 2026 AI Strategy Shift

On August 17, 2026, at the Alipay AI Ecosystem Conference, Ant Group CEO Han Xinyi made a call:

> "Agent commerce will explode within the next 6 to 12 months."

He cited three conditions coming together: improved AI capabilities, a leap in user demand, and the formation of value loops. He noted that while reach, production, and fulfillment efficiency have improved dramatically over the past few decades, information-conversion efficiency remains low — "traffic is running out, and traffic is too expensive" is a common merchant complaint.

Alipay's positioning is to "build the underlying layer" — providing model, trust, and connection capabilities without building front-end applications. Three unresolved problem sets remain: cross-device interconnection protocols, trust and payment mechanisms between agents, and a yet-to-take-shape revenue-sharing framework.

Ant Group's 2025 sustainability report disclosed total R&D spending of RMB 35.03 billion, up nearly 50% year over year. Han stated: "Ant is investing heavily in AGI frontier exploration and applied innovation, pushing AI deep into core business scenarios — Ant is back on the battlefield."

The August 28 launch of Ling-3.0-flash-Fin is a concrete move in this "back on the battlefield" strategy.

Ling-3.0-flash-Fin: A 124B / 5.1B Finance-Specific LLM

Released on August 28, 2026, the model is based on the Ling-3.0-flash architecture with 124B total parameters and 5.1B activated parameters, optimized for financial scenarios.

Capability improvements:

1. Continued pre-training on financial corpora — layering vertical financial data (IPO prospectuses, regulatory filings, brokerage research, listed-company announcements) on top of general corpora 2. Domain post-training — supervised fine-tuning for annual-report analysis, financial modeling, and research material processing 3. Tool-calling optimization — support for financial API calls (financial data, market data, macroeconomic databases)

Evaluation shows the model outperforms same-size models on FinFIRST and related benchmarks, while general capability improved to an AA Intelligence Index score of 41.

Per an August 28 DoNews report: "Ling-3.0-flash-Fin will offer one month of limited-time free API calls on OpenRouter, and model weights will be officially open-sourced next week."

Architecture Details

| Dimension | Value | | --- | --- | | Base architecture | Ling-3.0-flash (MoE) | | Total parameters | 124B | | Activated parameters | 5.1B | | Long context | Inherited from Ling-3.0-flash | | Training paradigm | Continued pre-training + domain post-training + tool-calling RL | | Benchmark results | Outperforms same-size on FinFIRST / AA Intelligence Index 41 | | Deployment | OpenRouter API + open-source weights next week |

> Tip: The 124B-total / 5.1B-activated MoE paradigm means actual inference cost is roughly 1/24 that of a dense 124B model. A month free on OpenRouter means any developer can trial the model at zero cost.

FalconTST 2.0: Time-Series Model for Cross-Border Payment FX Risk

On August 20, Ant International launched Falcon Time-Series Transformer 2.0, purpose-built for FX risk management in cross-border payments. It achieved state-of-the-art results on authoritative global benchmarks:

• MASE (Mean Absolute Scaled Error): 0.666
• Forecast accuracy: consistently above 93%

Major institutions including Barclays, Citi, Deutsche Bank, and Standard Chartered have applied FalconTST 2.0 to cash-flow forecasting and FX management:

• Barclays: integrated into its FX hedging platform BARX NetFX
• Citi: combined with its Fixed FX Rates solution
• Standard Chartered: integrated into the SCALE FX system
• Deutsche Bank + HSBC + 1 unnamed major bank

According to Kelvin Li (Li Yue), GM of Platform Technology at Ant International, Ant International has reached partnerships with 6 major banks. He also noted: "Precise forecasting can cut FX hedging and allocation costs by over 60%" — a very concrete commercial value proposition.

Time-Series vs. General: Two Roads in Financial AI

Per Ant International: "Unlike general-purpose LLMs, time-series models are better suited to financial and payment scenarios, parsing liquidity demand, FX volatility, and cash-flow changes in real time."

The underlying judgment: numerical prediction and text understanding in finance require different models.

• Text understanding (annual reports, research parsing, contract extraction): LLMs
• Numerical prediction (FX forecasting, liquidity forecasting, risk measurement): time-series models (TST)

Ant is betting on both roads — which is why Ling-3.0-flash-Fin (Aug 28) and FalconTST 2.0 (Aug 20) landed in the same month.

Global Comparison: Ant vs. Bloomberg GPT vs. FinGPT

| Dimension | Ling-3.0-flash-Fin | FalconTST 2.0 | Bloomberg GPT | FinGPT | | --- | --- | --- | --- | --- | | Model type | MoE LLM | Time-series Transformer | Closed-source LLM | Open-source LLM | | Size | 124B total / 5.1B active | Undisclosed | 50B | 7-13B | | Core use case | Annual reports / research / financial modeling | FX / liquidity / cash flow | Financial NLP | Financial NLP + sentiment | | Open source | Next week (Apache) | No | No | Yes | | Customers | 1-month OpenRouter free tier | 6 global banks | Bloomberg Terminal | Academia | | Benchmark | FinFIRST SOTA | MASE 0.666 SOTA | Internal benchmarks | Academic benchmarks |

Ant's dual-model strategy covers both core capabilities — text and numbers — something a single model like Bloomberg GPT does not.

Context: Brokers and Fund Managers Deploy AI (Aug 17-21)

• August 17: During the 10th annual 818 Wealth Festival, 11 securities brokers ran themed campaigns, with 7 making AI a core selling point — e.g., Soochow Securities' AI trio (signals + market monitoring + assistant), Guotai Haitong's Lingxi App 3.0, Sinolink's AI advisory covering pre-market selection through post-market review, and Western Securities' "AI Xiaoxi" agent in App 6.1.0.
• August 21: A survey of 20 leading fund companies showed a shift from "adding AI features" to "rethinking business logic around AI." E Fund, China Universal, China Southern, and Harvest have built enterprise-level AI platforms; China Universal has deployed DeepSeek, Qwen, and other mainstream open-source models. Most chose open-source models for localized deployment.

This resonates with Bailing's open-source decision — open-sourcing financial LLMs is a clear trend for H2 2026.

The Subtle Balance Between "Open Source" and "Finance"

• The financial industry is extremely sensitive to data privacy — clients won't process unpublished financials or trade data on closed models
• Open source + on-premises deployment is the key to adoption in finance
• But open source doesn't mean abandoning commercialization — Ant can attract developers via the OpenRouter free tier while serving global banks through FalconTST 2.0

My Take

Taken together, the two releases upgrade financial AI from "single-point capability" to a dual-model system: Ling-3.0-flash-Fin for text understanding, FalconTST 2.0 for numerical prediction.

Three questions worth watching skeptically:

1. Will open sourcing invite forks that bypass the business model? Will the license restrict use for training competing models? 2. Can FalconTST 2.0's promised 60% FX hedging cost reduction be sustained? It's a commercial claim needing 6-12 months of data. 3. Will "agent commerce exploding in 6-12 months" come true? Cross-device protocols, agent trust/payment, and revenue sharing remain unresolved.

Things to track over the next 6-12 months: GitHub stars / HuggingFace downloads after open-sourcing; FalconTST 2.0 expansion into credit, market, and operational risk; Ant's full "agent commerce" protocol stack; and how 7 brokers + 20 fund companies + Ant's dual models form a Chinese financial LLM ecosystem.

This is one of the most strategically significant Chinese financial AI events of H2 2026: a clear signal of the shift from single-point capability to dual-model systems + open source + global commercial deployment.