Ant Group AI Model Ling-1T: Trillion-parameter breakthrough

• Ant Group AI model Ling-1T scores 70% on elite math benchmark; dInfer framework allegedly 10x faster than Nvidia’s solution
• Open-source strategy aims to establish Chinese fintech giant as AGI infrastructure player

Chinese fintech giant Ant Group has significantly expanded its artificial intelligence (AI) capabilities with the launch of Ling-1T, a trillion-parameter language model, while simultaneously introducing an inference framework that the company claims outpaces solutions from chipmaking leader Nvidia by a factor of ten.

The Hangzhou-based company, operator of the Alipay mobile payment platform, announced the release and open-sourcing of the Ant Group AI model Ling-1T on October 9, positioning it as a flagship “non-thinking” model within its expanding Ling family.

The model demonstrates state-of-the-art performance on complex reasoning benchmarks while maintaining what the company describes as efficient inference capabilities.

According to Ant Group’s announcementLing-1T achieved a 70.42% accuracy rate on the 2025 American Invitational Mathematics Examination (AIME) benchmark at an average cost of over 4,000 output tokens per problem—performance the company states is comparable to best-in-class AI models currently available.

The model expansion follows Ant Group’s September release of Ring-1T-preview, which the company characterised as the world’s first open-source trillion-parameter “thinking model.”

Together, these releases signal an aggressive expansion of Ant Group’s AI model ecosystem, which now comprises three main series: the Ling non-thinking models, the Ring thinking models, and the multimodal Ming series, alongside an experimental model called LLaDA-MoE.

Speed claims challenge established players

Beyond the language model announcement, Ant Group introduced dInfer, an inference framework designed specifically for diffusion language models—a relatively new class of AI systems that generate outputs in parallel rather than sequentially.

The company claims dInfer operates up to three times faster than vLLM, an open-source inference engine developed by University of California, Berkeley researchersand ten times faster than Nvidia’s Fast-dLLM framework.

In internal testing on Ant Group’s diffusion model LLaDA-MoE, dInfer generated an average of 1,011 tokens per second on the HumanEval code-generation benchmark, compared with 91 tokens per second for Nvidia’s Fast-dLLM and 294 for Alibaba’s Qwen-2.5-3B model optimised with vLLM.

“We believe that dInfer provides both a practical toolkit and a standardised platform to accelerate research and development in the rapidly growing field of dLLMs,” Ant researchers wrote in a technical report accompanying the release.

Strategic positioning amid chip constraints

The aggressive push into alternative model paradigms and optimisation frameworks highlights how Chinese technology firms are leveraging algorithmic breakthroughs to compensate for disadvantages in accessing advanced AI chips.

While autoregressive language models—from OpenAI’s GPT-3.5 to DeepSeek’s R1—have powered the chatbot boom, companies like Ant Group continue exploring diffusion language models in pursuit of greater capabilities and efficiency.

He Zhengyu, Chief Technology Officer of Ant Group, framed the releases within a broader philosophical stance on AI development.

“At Ant Group, we believe Artificial General Intelligence (AGI) should be a public good—a shared milestone for humanity’s intelligent future,” He stated. “We are dedicated to building practical, inclusive AGI services that benefit everyone, which requires constantly pushing technology forward.”

Broader implications for AI competition

Ant Group’s developments arrive amid intensifying competition in China’s AI sector. Other major technology firms are similarly experimenting with alternative approaches—in July, ByteDance introduced Seed Diffusion Preview, claiming speeds five times faster than comparable autoregressive models.

The company is also developing AWorld, a development framework designed to support continual learning among AI agents that autonomously complete tasks for users, further expanding its AI infrastructure ambitions.

However, questions remain about the practical adoption of diffusion language models in production environments. While these systems show promise in specific benchmarks, autoregressive models continue dominating commercial applications due to their proven track record in understanding and generating human language.

The open-sourcing strategy Ant Group has adopted—making both Ling-1T and the dInfer framework publicly available—represents a calculated bet that collaborative development can accelerate innovation while establishing the company’s technologies as industry standards. Whether this approach can challenge established leaders in the AI space, particularly those with superior access to cutting-edge hardware, remains to be tested in real-world deployments.

For now, Ant Group’s rapid-fire releases signal that Chinese technology firms are refusing to cede ground in the global AI race, instead doubling down on software innovation and algorithmic efficiency as competitive differentiators.

Want to experience the full spectrum of enterprise technology innovation? Join TechEx in Amsterdam, California, and London. Covering AI, Big Data, Cyber Security, IoT, Digital Transformation, Intelligent Automation, Edge Computing, and Data Centres, TechEx brings together global leaders to share real-world use cases and in-depth insights. Click here for more information.