Ring-1T - Ant Financial's open-source trillion-parameter thinking model

What is Ring-1T?

Ring-1T is a trillion-parameter thinking model open-sourced by Ant Group. Based on the Ling 2.0 MoE architecture, it was pre-trained on a 20TB corpus and trained for reasoning capabilities using the self-developed reinforcement learning system ASystem. Supporting a 128k context window, it has performed close to or surpassed top-tier closed-source models in multiple international competitions and benchmark tests. It excels in natural language reasoning, scoring 92.6 in the AIME 2025 test, approaching GPT-5. In the IMO 2025 test, it solved the third question on its first attempt and provided partially correct answers to other questions, demonstrating high-order reasoning capabilities.

Main functions of Ring-1T

• Powerful natural language reasoning abilityAchieving a score of 92.6 in the AIME 2025 test, close to the GPT-5 score of 94.6, demonstrates strong mathematical reasoning ability.
• Highly efficient problem-solving capabilitiesIn the IMO 2025 test, the student solved question 3 on the first try and provided partial correct answers to the other questions, demonstrating high-order reasoning ability.
• Multi-domain competitivenessIt performs exceptionally well in tasks such as HMMT 2025, LiveCodeBench v6, CodeForces, and ARC-AGI-1, demonstrating broad applicability.
• Open source collaborationThe code and weights are completely open source and published on the Hugging Face platform to facilitate community exploration and feedback, and accelerate model iteration and improvement.

Ring-1T Technical Principles

• Architecture DesignIt adopts the Ling 2.0 MoE architecture, combined with trillions of parameters, to provide the model with powerful expressive power and high computational efficiency.
• Pre-training corpusPre-training was completed on 20T of high-quality corpus to ensure that the model can learn rich language knowledge and patterns.
• Reinforcement learning trainingWe improve the model's reasoning and decision-making abilities by training it with RLVR (Reinforcement Learning Reinforcement) based on its self-developed, high-efficiency reinforcement learning system, ASystem.
• Continuous iterationThe model is still being trained, continuously optimizing its performance and addressing current issues such as language mixing and repetitive inference.

Ring-1T project address

• Hugging Face Model Libraryhttps://huggingface.co/inclusionAI/Ring-1T-preview
• Ling Chathttps://ling.tbox.cn/chat

Ring-1T performance

• Mathematical reasoningIn the IMO 2025 test, Ring-1T solved problems 1, 3, 4 and 5 on the first try (silver medal level), and on the third attempt, provided a near-perfect proof for the geometric proof of problem 2.
• Code generationIt scored 2092 in Codeforces tests, surpassing GPT-5 (High)'s 2073; and ranked first among open-source models in programming benchmarks such as LiveCodeBench.
• Comprehensive abilityIt outperforms in tasks such as HealthBench and Creative Writing v3, with a win rate of 81.59% in the Arena-Hard-v2 test, approaching GPT-5's 82.91%.

Application scenarios of Ring-1T

• Natural Language Reasoning TaskAchieving a score of 92.6 on the AIME 25 test, close to the GPT-5 score of 94.6, demonstrates strong mathematical reasoning ability.
• Code generation and optimizationIt outperformed GPT-5 with a score of 94.69 in the CodeForces test, demonstrating its excellent code generation capabilities.
• Multi-agent framework applicationIt integrates with the multi-agent framework AWorld and can be used to test and explore complex reasoning tasks.
• Academic research and developmentAs the world's first open-source trillion-parameter inference model, it provides researchers and developers with a high-performance, reproducible inference foundation, promoting transparency and collaborative innovation in the large model ecosystem.