From the source
Ant Ling has extended Ling-3.0-flash into a financial domain model, Ling-3.0-flash-Fin, through continued training, task-specific post-training and tool-use optimization.
- The company says a general model can discuss a domain long before it can be trusted to work in one, and that professional tasks demand the right source, entity, reporting period and a checkable calculation.
- Ling-3.0-flash-Fin keeps the same sparse MoE architecture as Ling-3.0-flash, with 124B total parameters, 5.1B activated per inference step and a 256K context window.
- Its target is the chain of work around an answer: retrieve, verify, calculate, model, write and preserve enough context for the work to be inspected.
- Ant Ling says the model can read a company report and a seven-sheet Excel model with more than 5,000 formulas, reconcile a 2026 Q2 forecast with reported actuals and update linked formulas without breaking formula lineage.
- The company released FinFIRST alongside the model, an open benchmark for financial search agents developed with support from financial specialists.
- FinFIRST's first release contains 123 expert-authored and validated tasks from more than 50 financial professionals, with 701 atomic scoring items and 12,300 rubric points, and reports Atomic, Loose Pass and Strict Pass.
