China's group of antsa subsidiary of Alibaba, detailed technical details about its latest model, Ring-1Twhich, in accordance with the corporate, is “the primary open source reasoning model with a trillion total parameters.”
Ring-1T is meant to compete with other reasoning models comparable to GPT-5 and the O series OpenAIin addition to Google's Gemini 2.5. With the re-release of the most recent model, Ant is expanding the geopolitical debate about who will do it dominate the AI race: China or the USA.
According to Ant Group, Ring-1T is optimized for math and logic problems, code generation and scientific problem solving.
“With roughly 50 billion parameters enabled per token, Ring-1T achieves state-of-the-art performance in several demanding benchmarks – despite relying solely on natural language reasoning capabilities,” Ant said a paper.
Ring-1T, first released in preview in September, adopts the identical architecture as Ling 2.0 and is predicated on the bottom Ling-1T model that the corporate released earlier this month. Ant said this could allow the model to support as much as 128,000 tokens.
To train a model as large as Ring-1T, researchers needed to develop latest methods for scaling reinforcement learning (RL).
New training methods
Ant Group has developed three “connected innovations” to support Ring-1T’s RL and training, a challenge given the scale of the model and the typically high computational requirements related to it. These three are IcePop, C3PO++ and ASystem.
IcePop removes noisy gradient updates to stabilize training without slowing down inference. It helps eliminate a catastrophic misalignment of coaching and inference in RL. The researchers noted that when training models, especially those using a mixed-of-experts (MoE) architecture like Ring-1T, discrepancies in probability calculations can often occur.
“This problem is especially pronounced when training MoE models with RL attributable to the inherent use of the dynamic routing mechanism. Furthermore, in long CoT environments, these discrepancies can progressively accumulate and further amplify over iterations,” the researchers said.
IcePop “suppresses unstable training updates through double-sided masking calibration.”
The next latest method that the researchers needed to develop is C3PO++, an improved version of the C3PO system that Ant had previously established. The method manages how Ring-1T and other models with particularly large parameters generate and process training examples, or so-called rollouts, in order that GPUs don’t sit idle.
The way it really works would break the work into rollouts into parts that might be processed in parallel. One group is the inference pool, which generates latest data, and the opposite is the training pool, which collects results to update the model. C3PO++ creates a token budget to manage how much data is processed, ensuring GPUs are used efficiently.
The final latest method, ASystem, uses a SingleController+SPMD (Single Program, Multiple Data) architecture to enable asynchronous operations.
Benchmark results
Ant referred Ring-1T to benchmarks for measuring performance in math, coding, reasoning and general tasks. They tested it with models comparable to DeepSeek-V3.1-Terminus-Thinking, Qwen-35B-A22B-Thinking-2507, Gemini 2.5 Pro and GPT-5 Thinking.
In benchmark testing, Ring-1T performed strongly, rating second behind OpenAI's GPT-5 in most benchmarks. Ant said the Ring-1T performed the perfect of all open weight models tested.
The model scored 93.4% on the AIME 25 leaderboard, second only to GPT-5. When it got here to encoding, Ring-1T outperformed each DeepSeek and Qwen.
“It demonstrates that our rigorously synthesized data set shapes Ring-1T’s robust performance in programming applications, which provides a solid foundation for future efforts in agent applications,” the corporate said.
Ring-1T shows how much Chinese corporations are investing in models
Ring-1T is just the most recent model from China aiming to dethrone GPT-5 and Gemini.
Chinese corporations have been releasing impressive models at a rapid pace because the surprise launch of DeepSeek in January. Ant's parent company, Alibabarecently published Qwen3 Omnia multimodal model that natively combines text, images, audio and video. DeepSeek also continued to enhance its models earlier this month has launched DeepSeek OCR. This latest model reimagines the best way models process information.
As Ring-1T and Ant develop latest methods for training and scaling particularly large models, the battle for AI dominance between the US and China continues to accentuate.
