Open your banking app to dispute a charge, and the calm voice that talks you through it probably isn't a person. It's software — and until recently, the company behind it would have kept that software under lock and key. Ant Group just did the opposite.
In October 2025, Ant released and open-sourced Ling-1T, a general-purpose large language model (大模型) with roughly one trillion total parameters. It is the flagship of the company's "Ling" (百灵) family, which also spans a "thinking" series called Ring and a multimodal series called Ming. The release is notable less for its size than for what Ant intends to do with it: put an open-weight model into the regulated, money-moving products that hundreds of millions of people already use.
Key takeaways
- Ant Group released and open-sourced Ling-1T, a roughly one-trillion-parameter general-purpose language model, in October 2025.
- The model uses a Mixture-of-Experts (MoE) design that activates only a small slice of its weights per query, and was trained with FP8 mixed precision.
- Ant is embedding it inside Alipay's AI health, finance, and lifestyle assistants instead of selling it as a standalone product.
- A companion reasoning model, Ring-1T-preview, was open-sourced in September 2025 as what Ant describes as the first open-source trillion-parameter "thinking" model.
- The full Ling family — Ling, Ring, Ming, plus an experimental LLaDA-MoE — is downloadable from Hugging Face and ModelScope (魔搭).
What Ling-1T actually is
Ling-1T is a non-thinking model: it answers directly rather than producing a long internal chain of reasoning before replying. Ant says the design targets a balance between precise reasoning and efficient inference, with strong results on code generation, software development, competition-level mathematics, and logical reasoning.
Technically, it is a high-sparsity, fine-grained MoE model. Ant's disclosures state a total of one trillion parameters spread across 256 expert blocks, with only about 50 billion parameters (roughly 3.5%) activated for each token. It was pre-trained on more than 20 trillion tokens of high-reasoning-density data and supports a context window of up to 128,000 tokens.
The engineering choices that matter
Two details stand out for practitioners. First, Ling-1T was trained with FP8 mixed precision across its full run — Ant describes it as among the largest foundation models (基础模型) known to use FP8 throughout training, which the team credits with meaningful memory savings and end-to-end speedups. Second, Ant published a sentence-grain policy-optimization method it calls LingPO, intended to keep training stable at trillion-parameter scale.
Those are the kinds of choices that rarely make headlines but decide whether a model can be served cheaply at scale. A trillion-parameter dense model would be impractical for most deployments; an MoE that fires only about 5% of its weights per token is a different economic proposition, and it is the reason a payments company can justify running these models inside consumer products.
Why a payments company gives models away
Ant operates Alipay, China's dominant mobile-payment platform. Its incentive is not to sell model access but to make its own products smarter and cheaper to run. The Ling family already powers Ant's AI health assistant, AI finance assistant, and AI lifestyle assistant, and the company opens model capabilities through Alipay's "百宝箱" (treasure-box) platform.
According to Li An, chief scientist at BrainNet (脑机网), the more interesting move is not the parameter count but Ant's willingness to embed open-weight models directly into regulated financial workflows, where most Western labs hesitate to deploy.
Where it shows up in real products
For an ordinary user, the difference is felt inside Alipay: a health assistant that can summarize a medical report, a finance assistant that explains a transaction, a lifestyle assistant that books or compares services. Ant's pitch is "trustworthy intelligence for industry" — keeping sensitive data and decision flows inside its own stack while still publishing the underlying models.
The open-source releases also include Ant's own specialist models for structured data, mathematics, and code, plus evaluation sets, available on platforms including Hugging Face, ModelScope, and Gitee. For developers, that means a trillion-parameter-class model can be pulled, fine-tuned, and self-hosted rather than rented through an API.
Honest limitations
This article relies on Ant Group's own press materials, technical posts, and Chinese financial-news coverage; we did not independently benchmark the model. Headline benchmark figures quoted by Ant, such as AIME 2025 results, come from the company and were not cross-checked against an independent lab. The "first open-source trillion-parameter thinking model" claim is Ant's own framing. Real-world latency, cost, and safety behavior in production depend on Ant's serving stack, which is not public. We also did not test multilingual or non-finance performance, and open-source availability may differ by region or license.
What readers can do now
- If you build applications, pull Ling-1T or Ring-1T-preview from Hugging Face or ModelScope and benchmark them on your own tasks before trusting vendor numbers.
- If you use Alipay, try the health, finance, and lifestyle assistants and notice where responses are genuinely useful versus where they defer to human review.
- If you follow open-weight trends, track whether Ant keeps releasing larger Ring and Ming variants — that signals how serious the company is about open foundation models versus marketing.