When you tap to pay with Alipay, you are not thinking about large language models. Ant Group, the company behind that wallet, has been thinking about them a lot — and this autumn it dropped a trillion-parameter model into the open-source pool.
The model is Ling-1T. It is the flagship of Ling 2.0, a family Ant Group began publishing on 10 September 2025. Unlike a research toy, it is pitched straight at the boring, high-stakes work Ant actually does: reading documents, reasoning about money, and serving enterprises that cannot afford hallucinations.
The headline number
Ling-1T is large:
- Total parameters: about 1 trillion (1T).
- Activated per token: roughly 50 billion.
- Context window: up to 128K tokens.
- Pre-training corpus: more than 20 trillion tokens.
So, like Meituan's LongCat, the trick is not raw size but how little of the model runs at once. Ant says it wakes about 5% of the brain per token.
A family, not a single release
Ling 2.0 shipped in tiers, all under the MoE (mixture-of-experts) design with a 1/32 activation ratio:
- Ling-mini-2.0 — 16B total, 1.4B activated, 300+ tokens/s on H20.
- Ling-flash-2.0 — 100B total, 6.1B activated, 200+ tokens/s on H20.
- Ling-1T — the 1T flagship described above.
The small ones are the interesting part for most teams: a 1.4B-activated model that claims to punch at the level of a 7–8B dense model is cheap to serve.
Why efficiency is the whole point
Ant's stated thesis is "efficient reasoning" — get the right answer with fewer tokens and less silicon. Two engineering choices back it:
- FP8 mixed-precision training, which Ant says cut memory use and lifted end-to-end speed by 15% or more.
- A "mid-training + post-training" pipeline plus an Evolutionary Chain-of-Thought (Evo-CoT) method to make the model think economically.
On ArtifactsBench, a frontend-coding benchmark, Ant claims Ling-1T ranked first among open models. We treat that as the company's own claim until independent labs confirm it.
Where this clearly points: finance
Ant is not a general-purpose lab competing for chatbot mindshare. Its natural customers are banks, insurers, and the millions of small merchants already inside Alipay's ecosystem. A large model (大模型) that reasons about transactions, flags risk, and explains itself in plain language is a product Ant can sell without leaving its own walled garden.
The open-source release is also a recruiting and credibility play: publish weights, let researchers poke holes, and the model improves faster than a closed one would.
The open-source calculus
Ant open-sourced through Hugging Face and GitHub, including the FP8 training recipe — a move that helps smaller labs fine-tune on limited hardware. That lowers the barrier for exactly the kind of enterprise partners Ant wants.
Why a payments company builds models
Ant's logic is vertical integration. It already owns the distribution (Alipay), the data (transactions), and the compliance relationships (banks). A model that lives inside that stack turns a cost center — customer support, risk review — into a margin lever. Open-sourcing the research-grade parts buys outside stress-testing without giving away the proprietary data that makes the product defensible.
The risk Ant is taking
Giving away a trillion-parameter flagship is not free. The strongest argument against open-sourcing Ling-1T is that it hands a capable base model to competitors who can fine-tune it for the same finance and risk jobs Ant wants to own. Ant's answer is the oldest one in enterprise software: the moat is the data and the compliance pipe, not the model weights. A bank is not going to swap Alipay's regulated rails for a downloaded checkpoint. So Ant trades a little model exclusivity for a larger ecosystem that validates and extends Ling faster than a closed team could.
What the 1T size actually buys
Scale is not vanity here if the activation stays low. A 1T model that wakes only 50B per token can absorb more training signal and more languages than a 100B model, while serving at a cost closer to that smaller model. The bet is that, for hard reasoning and long documents, the headroom of a trillion parameters shows up as fewer wrong answers — the thing finance customers actually pay for. Whether it does is exactly what the open release will now let outsiders measure.
Honest limitations
- The trillion-parameter figure and benchmark rankings come from Ant's own inclusionAI blog and announcements; we did not reproduce them.
- "First among open models" on ArtifactsBench is Ant's framing; independent comparison is still thin.
- FP8 speedup and token/s numbers are vendor-measured on specific hardware (H20), not on commodity GPUs most readers own.
- We have not verified the 20T-token corpus or the 128K context in practice; context length on real long documents can degrade.
- Open weights are not open data; the training set is not public.
What readers can do now
- Pull Ling-mini-2.0 or Ling-flash-2.0 from Hugging Face and run a finance or document task you already do.
- If you serve Chinese enterprise users, test Ling against Qwen and DeepSeek on your own data, not leaderboards.
- Watch the FP8 training recipe — it is the part most useful to small teams with tight GPUs.
- Don't assume bigger is better here; the 1.4B-activated mini model may beat the 1T flagship on your latency budget.
- If you need a model that "explains money," Ling is worth a serious pilot; if you need casual chat, almost anything smaller will do.