Ant Group Launches Ling-1T: A Trillion-Parameter AI Model

Ant Group has officially thrown its hat into the ring of trillion-parameter AI models with the open-sourcing of Ling-1T. The Chinese fintech giant is positioning this language model as a critical advancement, claiming it strikes a pivotal balance between robust computational efficiency and sophisticated reasoning capabilities, a key challenge in the current AI landscape.

The announcement, made recently, signals a major step for Ant Group, the operator of Alipay. The company has been aggressively investing in building a comprehensive artificial intelligence infrastructure encompassing various model architectures designed to enhance their offerings.

Ling-1T demonstrates competitive performance in handling complex mathematical reasoning, a crucial test for advanced AI systems. The model achieved an accuracy rate of 70.42% on the 2025 American Invitational Mathematics Examination (AIME) benchmark. This testing assesses the ability of the AI model to understand, process and solve complex mathematical thinking with logical accuracy.

According to Ant Group’s technical specifications, Ling-1T maintains this performance level while processing an average of over 4,000 output tokens per problem. This positions it competitively amongst the “best-in-class AI models” in terms of high-quality results. Analysts suggest that this feat is attributable to a novel architectural optimisation.

Dual-pronged approach to AI advancement

The release of the trillion-parameter AI model is strategically aligned with the launch of dInfer, a specialized inference framework tailored for diffusion language models. This dual release underscores Ant Group’s pragmatic strategy of exploring and investing across multiple technological approaches, rather than committing to a single AI architectural paradigm. dInfer is designed to accelerate the development and deployment of diffusion language models, a technology that has the potential to unlock new frontiers in AI.

Diffusion language models represent a significant shift from the more prevalent autoregressive systems that power widely used chatbots like ChatGPT. Unlike the sequential text generation of autoregressive models, diffusion models generate outputs in parallel. This approach, already widely used in image and video generation tools, is less common but increasingly viable in language processing.

Ant Group’s testing metrics indicate a notable efficiency boost with dInfer. When tested on the company’s LLaDA-MoE diffusion model, dInfer yielded 1,011 tokens per second on the HumanEval coding benchmark, significantly outperforming NVIDIA’s Fast-dLLM framework (91 tokens per second) and Alibaba’s Qwen-2.5-3B model running on vLLM infrastructure (294 tokens per second). These gains could transform how developers approach building AI-powered applications.

“We believe that dInfer provides a practical toolkit and a standardized platform to accelerate research and development in the rapidly growing field of dLLMs,” Ant Group researchers stated in the technical documentation. This highlights the strategic importance of providing tools that democratize access to these emerging technologies.

Ecosystem expansion beyond language models

The Ling-1T trillion-parameter AI model is a component within a broader ecosystem of AI systems that Ant Group has strategically built in recent months, representing a deliberate and comprehensive expansion into the artificial intelligence industry.

Ant Group’s AI portfolio now encompasses three primary series: the Ling series of “non-thinking” models for standard language tasks, the Ring series of more sophisticated “thinking” models designed for complex reasoning (including the previously released Ring-1T-preview), and the Ming series of multimodal models capable of processing input data across images, text, audio, and video. This diversified approach reduces the risk inherent in relying on any single technology or model architecture.

These varied technologies extend to an experimental model known as LLaDA-MoE, which leverages Mixture-of-Experts (MoE) architecture. MoE selectively activates the most relevant portions of a massive model for specific tasks—a technology that promises significant efficiency gains. The MoE approach is important for increasing model efficiency, reducing computational cost and allowing for the building of more sophisticated systems.

He Zhengyu, chief technology officer at Ant Group, has articulated the company’s vision for these new technological releases. “At Ant Group, we believe Artificial General Intelligence (AGI) should be a public good—a shared milestone for humanity’s intelligent future,” He stated. He characterized the open-sourcing of the trillion-parameter AI model and Ring-1T-preview as initial steps towards “open and collaborative advancement,” highlighting the philosophical importance of these new AI solutions.

Competitive dynamics in a constrained environment

The carefully planned release timings and characteristics of Ant Group’s releases reveal the strategic calculations being made within China’s AI sector. With access to cutting-edge semiconductor technology increasingly limited by export restrictions, Chinese technology firms are placing greater emphasis on algorithmic innovation and software optimisation to build competitive advantages.

ByteDance, the parent company of TikTok, introduced a related diffusion language model, Seed Diffusion Preview, in July, advertising speed improvements of up to five-fold over comparable autoregressive architectures. These simultaneous innovative efforts across companies, illustrate the industry’s shared pursuit of model paradigms that provide efficiency benefits.

Regardless, the path to widespread practical adoption for diffusion language models has yet to be definitively laid out. Autoregressive systems remain dominant in commercial deployments due to their demonstrable performance in natural language understanding and generation—the fundamental requirements for many of the industry’s most popular customer-facing applications.

Open-source strategy as market positioning

By making the trillion-parameter AI model open-source, alongside the dInfer framework, Ant Group is encouraging a collaborative development model that sets them apart from the closed approaches of some of its biggest competitors. Open-sourcing provides a method by which Ant Group can collaborate with AI developers, and in doing so the company hopes to stay at the cutting edge of AI innovation.

This approach carries the potential to quickly accelerate innovation while strategically positioning Ant Group’s technologies as the core infrastructure for the broader AI community. Collaboration is a key factor in the expansion of new technologies through adoption and standardisation.

The company is also simultaneously cultivating AWorld, a novel framework intended to support consistent learning in autonomous AI agents – systems designed to perform complex automation tasks independently on behalf of users. Ant Group’s investments extend beyond just the models into systems and technologies for automation.

Whether these coordinated efforts can firmly establish Ant Group as a leading power in global AI development depends partly on the validation of performance claims by users in real-world deployments. Equally important is the adoption rate among developers seeking alternatives to established platforms. The adoption of these new technologies will depend on how well they perform in real-world applications.

The open-source nature of the AI model may accelerate the technical validation process, while simultaneously building a critical mass of users. Building a strong user community will be important in testing and identifying new uses cases.

At this point, Ant Group’s releases demonstrate that major Chinese technology firms view the AI landscape as fluid enough to sustain new entrants willing to innovate across multiple dimensions.