Ant Group is entering the trillion-parameter model race with an unusually broad wager. Its newly open-sourced Ling-1T language model pursues scale and mathematical reasoning, while a separate inference framework called dInfer explores a different way to generate language at much higher speed.
Ling-1T achieved 70.42 percent accuracy on the 2025 American Invitational Mathematics Examination benchmark, according to Ant Group. It used an average of more than 4,000 output tokens for each problem, placing the model in what the company describes as best-in-class territory for result quality.
The release matters beyond a single score. Ant is simultaneously developing conventional models, reasoning systems, multimodal AI, diffusion language technology, and infrastructure for autonomous agents. Rather than betting that one architecture will settle the AI race, the Alipay operator is building several paths and inviting outside developers to test them.
One Portfolio Serves Several Kinds of Intelligence
Ling-1T belongs to a family divided by purpose. The Ling series handles standard language tasks without an explicit “thinking” mode. Ring models, including Ring-1T-preview, target complex reasoning. Ming systems process combinations of images, text, audio, and video.
An experimental model named LLaDA-MoE adds two more ideas. It uses a Mixture-of-Experts architecture, activating only the portions needed for a particular task, and operates as a diffusion language model. Both choices seek efficiency: a very large model need not engage all of its capacity for every request, while output generation need not always proceed one token at a time.
Ant Group chief technology officer He Zhengyu argues that artificial general intelligence should be a public good and a shared milestone. Open-sourcing Ling-1T and Ring-1T-preview gives researchers and developers a chance to examine those claims, reproduce results, and discover where the models succeed or fail outside Ant’s own tests.
Diffusion Challenges the One-Token-at-a-Time Model
Most widely deployed chatbots are autoregressive. They generate text sequentially, predicting the next token from everything that came before it. Diffusion models instead work on outputs in parallel, adapting a technique more familiar from image and video generation to language.
Ant’s dInfer framework is designed to make that approach practical. On the HumanEval coding benchmark, the company reports that LLaDA-MoE running with dInfer produced 1,011 tokens per second. NVIDIA’s Fast-dLLM framework reached 91, while Alibaba’s Qwen-2.5-3B on vLLM reached 294 in the cited comparison.
Those figures are striking, but speed is not the whole user experience. Autoregressive systems retain an advantage in commercial adoption because their language performance is proven across many applications. Diffusion models must show that parallel generation remains coherent, controllable, and useful across real workloads, not only that it moves quickly through a benchmark.
Open Access Turns Performance Claims into a Test
China’s AI companies face restricted access to cutting-edge semiconductors, increasing the strategic importance of algorithm design and software optimisation. ByteDance has also introduced a diffusion language model, Seed Diffusion Preview, with a claimed fivefold speed improvement. The parallel effort suggests that alternative generation methods are becoming a meaningful research direction rather than a curiosity.
Ant is also developing AWorld, a framework for continuous learning in autonomous agents. Together, the projects span models, inference, multimodal interaction, and systems that act for users. The breadth could produce a connected platform, although it also creates the challenge of turning many experiments into technology developers consistently adopt.
Open sourcing makes that outcome easier to judge. Ling-1T’s benchmark result and dInfer’s throughput can now meet independent hardware, unfamiliar tasks, and practical constraints. If a community forms around the releases, Ant’s tools could become infrastructure beyond its own products. If it does not, the public code will still help reveal whether trillion-parameter scale and diffusion generation offer advantages that survive outside the laboratory.
