Recently, Ant Group announced the open-source release of its self-developed trillion-parameter inference large model, Ring-1T-preview, which has become the world's first open-source trillion-parameter inference model. Although it is currently only a preview version, Ring-1T-preview shows excellent performance in natural language reasoning, with multiple test results demonstrating impressive capabilities.
According to official information, Ring-1T-preview scored 92.6 on the AIME25 test, surpassing all known open-source models, including Gemini2.5Pro, and approaching GPT-5's score of 94.6. In the CodeForces test, Ring-1T-preview achieved a high score of 94.69, demonstrating strong code generation capabilities and leading over GPT-5. This achievement also places it at the top of authoritative rankings such as LiveCodeBench and ARC-AGI-v1 among open-source models.
The Ant Bailing team also tested the reasoning capabilities of Ring-1T-preview in the International Mathematical Olympiad (IMO25). The test results showed that the model could correctly solve the third question in one go and derive partial correct answers for the first, second, fourth, and fifth questions. This indicates that Ring-1T-preview also performs well in complex reasoning tasks.
Currently, the Ant Bailing team is conducting post-training on the Ling2.0 family 1T language base, aiming to further explore the potential of this trillion-parameter model in natural language reasoning. At the same time, the official version of Ring-1T is also under active training and is expected to be released in the future.
This new development marks an important breakthrough in the field of artificial intelligence, especially in the development of open-source models. By opening the source code, Ant Group hopes to provide more possibilities for the research and application of AI technology, promoting the common progress of the industry.
Key Points:
🌟 Ant Group has launched the world's first open-source trillion-parameter inference large model, Ring-1T-preview.
🚀 Ring-1T-preview performs excellently in multiple tests, surpassing several known open-source models and approaching GPT-5.
🧠 The team is conducting follow-up training on the model to maximize its natural language reasoning capabilities.
