Model Card｜Ling-3.0-flash-Fin

🌏 中文版

Model Information

Key Capabilities

• Finance-domain specialization: Continued pretraining on high-quality financial data and domain-specific post-training on top of Ling-3.0-flash, optimized for annual reports, financial workbooks, and multi-document research materials while maintaining deployment efficiency
• End-to-end research pipeline: Connects information retrieval → evidence review → calculation → modeling → report generation as a unified workflow rather than isolated tasks
• Source-grounded financial search: Prioritizes authoritative primary sources for accurate, complete, and traceable answers; the FinFIRST benchmark is open-sourced alongside the model to enable transparent evaluation
• Multi-document financial reasoning: Reconciles reporting periods, definitions, assumptions, and conflicting figures across annual reports, earnings releases, regulatory filings, and research materials
• Valuation and spreadsheet workflows: Understands formulas, actual-vs-estimate updates, cross-sheet dependencies, balance checks, scenario analysis, and delivers editable financial models
• AA Intelligence Index v4.1.1: Improved from 38 (general-purpose version) to 41, demonstrating that domain enhancement does not come at the cost of general capabilities

Benchmark Performance

Ling-3.0-flash-Fin was evaluated across the following financial agent benchmarks. Official results are described qualitatively as "competitive with similarly sized models and substantially larger general-purpose models, with particular strength in source selection and tool-intensive financial tasks." FinFIRST will be open-sourced shortly to provide a transparent evaluation standard.

⚠️ The only publicly quantified metric is the AA Intelligence Index v4.1.1 score of 41 (up from 38 in the general Ling-3.0-flash). Full benchmark score tables are not yet released by Ant Group.

Comparison Within the Ling 3.0 Family

Compared to the base Ling-3.0-flash, flash-Fin shares the same architecture and context capabilities; the main difference is financial domain adaptation. For general-purpose tasks (coding, search, general reasoning), Ling-3.0-flash remains the better fit.

Architecture Details

Ling-3.0-flash-Fin retains Ling-3.0-flash's core architecture:

• Native hybrid linear attention: KDA (Kimi Delta Attention) and MLA layers alternate at a 5:1 ratio. KDA evolved from Lightning Attention with fine-grained diagonal gating in Delta Rule state updates, retaining critical information more precisely when processing long documents and extensive codebases
• 1/64 sparse MoE: Expert activation ratio compressed from 1/32 in the previous generation, yielding significantly higher "efficiency leverage"
• 256K → 1M context: Natively supports 256K, seamlessly scalable to 1M tokens
• TTFT optimization: Cluster-level hierarchical caching (SGLang HiCache + Mooncake) eliminates redundant recomputation, reducing Time-to-First-Token by 60–80% in long-input scenarios

Implications for Agent Development

If you are building agents for financial research, investment analysis, earnings interpretation, or valuation modeling: Ling-3.0-flash-Fin is currently one of the best open-source options available — the MIT license means you can self-host, the 124B/5.1B profile makes it potentially runnable on consumer hardware (see the Edge0 framework running a 35B MoE on Mac mini M4 Pro with only 2.9 GiB), and OpenRouter's free tier lets you validate quickly without commitment.

If you are building general-purpose agents (coding, search, general reasoning): Ling-3.0-flash or Ling-3.0-tiny may be more appropriate — the financial domain fine-tuning advantages of flash-Fin won't materialize in non-financial scenarios.

Not recommended for: Production financial decision-making requiring 100% certainty — the official documentation explicitly states that "key assumptions, valuation results, and investment conclusions require professional review and do not constitute investment advice." As the first finance-enhanced release, it still requires further validation in complex long-horizon workflows.

Takeaways

What stands out here is the "finance-enhanced model" category being opened up properly. Rather than slapping a financial API on top of a general model, Ant Group released the full model weights (MIT) alongside a dedicated evaluation benchmark (FinFIRST). This is consistent with the Ling 3.0 family's "planning-execution separation" philosophy: Ling-3.0-flash handles high-speed execution, and flash-Fin layers domain depth on top. The key thing to watch is whether FinFIRST becomes a de facto standard in financial AI once it is open-sourced.

References

• Ant Group Official Press Release: Open-Sources Ling-3.0-flash-Fin for Real-World Financial Workflows
• Hugging Face: inclusionAI/Ling-3.0-flash-Fin
• ModelScope: Ling-3.0-flash-Fin
• BusinessWire: Ant Group Unveils Ling-3.0-Flash Delivering Top-Tier Performance at a Fraction of the Parameter Scale (2026-07-27)
• FinFIRST (Hugging Face Dataset)
• OpenRouter: Ling-3.0-flash-Fin:free
• developer.ant-ling.com: Ling-3.0-flash Release Announcement
• IT之家: 蚂蚁百灵推出金融增强模型 Ling-3.0-flash-Fin