Ant Group's Ling (百灵) team put Ling-3.0-flash-Fin live at 15:58:10 UTC on 27 August, timestamped by the creation record on the routing layer serving it. The Chinese-language announcement followed at 02:41:40 UTC on 28 August. The model is a mixture-of-experts tuned on financial corpora: 124 billion total parameters, 5.1 billion active.
What actually shipped
One serving endpoint, whose internal id carries the vendor's own date stamp — inclusionai/ling-3.0-flash-fin-20260827:free — with prompt and completion both priced at zero for a stated one-month window, through a single provider. Context length is 262,144 tokens with a maximum completion of 32,768.
What the common framing gets wrong
English coverage is calling this an open-source release. It is not, today. The weights are unreleased; Ant says "next week," and its inclusionAI Hugging Face organisation currently lists no Ling-3.0-flash-Fin repository at all — the newest entries there are UI-Venus-2-9B and Ling-3.0-flash base variants from eight days ago. Open weights here is a promise with a rough date, not an event that has happened.
Second, the benchmark numbers are self-reported. The rise from 38 to 41 on an intelligence index appears in Ant's own announcement material, not in an independent publication by the index's maintainers. The finance-agent sweep — FinFIRST, FinSearchComp Verified, Finance Agent, APEX-Agents, SpreadsheetBench, τ³-Banking — is likewise Ant's own, and FinFIRST is a benchmark Ant describes designing with 50-plus financial professionals it recruited. A vendor scoring well on a vendor-built benchmark is a weaker claim than the number suggests.
Third, the parameter counts are not new. 124B total and 5.1B active are inherited unchanged from Ling-3.0-flash, released on 24 July. This is a continued-pretrain plus post-train on financial data, not a new frontier model, and coverage treating it as one inflates it. Smaller point: the context gets reported as "256K", but the served value is 262,144 with a 32,768-token output ceiling — a real constraint on the long-report use case the model is being sold for.