Ant's finance model is free to use — and the only one in the family you cannot download

Every Ling-3.0 model InclusionAI published this summer landed the same way: weights on Hugging Face, a licence line in the model card, usually MIT. The finance variant that appeared at hosting providers on 27 August 2026 broke the pattern, and a week later the gap has not closed. Ling-3.0-flash-fin is served as an endpoint at OpenRouter, Novita AI and Vercel AI Gateway. It is not in the inclusionAI organisation on Hugging Face, not on ModelScope, and not mentioned in the publisher's own repositories or in its deployment cookbook, which lists build after build of the base model. Checked on 1 September 2026, the organisation's newest uploads are a GUI agent and quantised builds of Ling-3.0-tiny — no finance model among them. The specifications quoted by the providers are inherited from the base model: 124 billion parameters with roughly 5.1 billion active per token, a 262,144-token context and 32,768 tokens of output. Nobody has published a benchmark, a training-data description or an evaluation on financial tasks. The claim that the model is better at investment work than its parent currently rests on a vendor blurb. The price is the other half of the story. OpenRouter carries only a zero-priced listing, and Novita charges nothing for input or output — while the ordinary Ling-3.0-flash on the same platform is billed. Free access with no weights and no paper is the shape of a public trial, not of a product, and such terms can be withdrawn at any time. For a reader in Europe the practical difference is sharp. The open Ling models can be downloaded, inspected and run on your own hardware, including on financial documents that never leave the building. This one can only be reached by sending those documents to a Chinese provider's server. That is a different decision entirely, whatever the invoice says. wujec.ai lists the variant as a separate profile from Ling-3.0-flash, precisely because of that difference.