124B total / 5.1B active (MoE; 127.49B stored parameters)
262,144 tokens
MIT
The first finance-enhanced model in the Ant Ling family, developed by Ant Group with financial institutions and domain experts as a continued-training derivative of Ling-3.0-flash. Ling-3.0-flash-Fin retains the base model's 42-layer BailingMoeV3 hybrid-linear Mixture-of-Experts architecture, comprising 35 Kimi Delta Attention layers and seven gated Multi-head Latent Attention layers in a repeating 5:1 pattern. It has two initial dense layers, 512 routed experts plus one shared expert, activates eight routed experts per token, and uses a 2,560-dimensional hidden state with 32 attention heads. The model is marketed as containing 124B total and 5.1B active parameters, while its Hugging Face safetensors metadata counts 127.49B stored parameters. Its configuration supports a 262,144-token context window. inclusionAI continued training the model on high-quality financial data to support end-to-end financial research workflows that combine information retrieval, source review, calculation, modeling, and report preparation. Targeted capabilities include reconciling figures, definitions, assumptions, and reporting periods across annual reports, earnings releases, regulatory filings, and research documents; source-grounded financial search; valuation modeling; spreadsheet formulas and cross-sheet dependencies; actual-versus-estimate updates; balance checks; scenario analysis; and production of editable research materials. The accompanying FinFIRST benchmark contains 123 bilingual Chinese and English financial-search problems with detailed rubrics, source requirements, calculations, and temporal constraints. In inclusionAI's reported evaluations, Ling-3.0-flash-Fin scores 52.85 on FinFIRST, 77.04 on FinSearchComp Verified, 51.74 on FinCRAFT, 69.19 on Finance Agent v1.1, 59.81 on Finance Agent v2, 29.17 on APEX-Agents, 86.50 on SpreadsheetBench V1, 21.81 on SpreadsheetBench V2, and 41.00 on tau3-Banking. The evaluations use different benchmark-specific harnesses: financial-search tests use ReAct with web search, page visits, and Python; SpreadsheetBench uses a Claude Code harness with long turn and time limits; and the tau3-Banking result is sourced from Artificial Analysis. Thinking is enabled by default, the chat template supports disabling it and structured tool calls, and inclusionAI recommends temperature 1.0, top-p 0.95, and top-k 20 for general inference. The initial open checkpoint is released only in BF16 as 64 main weight shards plus an MTP checkpoint and is compatible with the same SGLang and Ling-specific vLLM runtimes as Ling-3.0-flash. inclusionAI cautions that financial assumptions, valuations, and investment conclusions require professional review and says future releases will explore finance-enhanced models at larger scales.