from china ant groupAn affiliate of Alibaba, detailed technical information about its new model, Ring-1TWhat the company calls “the first open-source reasoning model with one trillion total parameters.”
The Ring-1T aims to compete with other reasoning models such as the GPT-5 and O-series OpenAIas well as GoogleGemini 2.5. With new release of latest model, Ant expands geopolitical debate over who will Dominate the AI race: China or America.
Ant Group said Ring-1T is optimized for mathematical and logical problems, code generation and scientific problem-solving.
“With approximately 50 billion active parameters per token, Ring-1T achieves state-of-the-art performance across multiple challenging benchmarks – despite relying entirely on natural language reasoning capabilities,” said Ant. a paper,
Ring-1T, which was first released on preview in September, adopts the same architecture as Ring 2.0 and is trained on the Ring-1T-base model that the company released earlier this month. Ant said this allows the model to support up to 128,000 tokens.
To train large models like Ring-1T, researchers had to develop new methods to enhance reinforcement learning (RL).
new methods of training
The Ant Group developed three “interconnected innovations” to support RL and training of Ring-1T, a challenge given the model’s size and generally large computation requirements. These three are Icepop, C3PO++ and ASystem.
Icepop removes noisy gradient updates to stabilize training without slowing down inference. This helps eliminate destructive training-prediction misalignment in RL. The researchers noted that when training models, especially using mixture-of-experts (MOE) architectures such as Ring-1T, there can often be inconsistency in probability calculations.
“This problem is particularly pronounced in training MOE models with RL due to the implicit use of dynamic routing mechanisms. Additionally, in long COT settings, these inconsistencies can gradually accumulate and propagate across iterations,” the researchers said.
Icepop “suppresses unstable training updates through two-way masking calibration.”
The researchers next had to develop the new method C3PO++, an improved version of the C3PO system previously established by Ant. The method manages how Ring-1T and other extra-large parameter models generate and process training examples, or what they call rollout, so that GPUs don’t sit idle.
The way it works is it will break the work into pieces in the rollout to process in parallel. One group is the inference pool, which generates new data, and the other is the training pool, which collects results to update the model. C3PO++ creates a token budget to control how much data is processed, ensuring that the GPU is used efficiently.
The final new method, ASystem, adopts a single controller + SPMD (single program, multiple data) architecture to enable asynchronous operation.
benchmark results
Ant pointed the Ring-1T to benchmarks measuring performance in math, coding, logical reasoning, and general tasks. They tested it against models like DeepSeq-v3.1-terminus-thinking, QUEN-35b-a22b-thinking-2507, Gemini 2.5 Pro, and GPT-5 thinking.
In benchmark testing, Ring-1T performed strongly and ranked second behind OpenAI’s GPT-5 in most benchmarks. Ant said the Ring-1T performed the best among all the open-weight models tested.
The model posted a score of 93.4% on the AIME 25 leaderboard, second only to GPT-5. In coding, Ring-1T outperformed both DeepSeek and Quen.
“This indicates that our carefully synthesized dataset shapes Ring-1T’s strong performance on programming applications, creating a strong foundation for future efforts on agentic applications,” the company said.
Ring-1T shows how much Chinese companies are investing in the models
The Ring-1T is China’s latest model which aims to dethrone the GPT-5 and Gemini.
Since DeepSeek’s surprise launch in January, Chinese companies have been releasing impressive models at a rapid pace. Ant’s parent company, alibabarecently released QUEN3-OmniA multimodal model that seamlessly integrates text, image, audio and video. DeepSeek has also continued to improve its models and earlier this month, DeepSeek-OCR launchedThis new model reimagines how models process information.
The battle for AI dominance between the US and China continues to escalate, with Ring-1T and Ant developing new ways to train and scale extra-large models.
