What is Ring-1T
Ring-1T is a trillion-parameter thinking large model open-sourced by Ant Group, based on the Ling 2.0 MoE architecture, pretrained on 20T of corpus, and trained for reasoning capabilities through the self-developed reinforcement learning system ASystem. It supports a 128k context window and performs close to or surpasses top closed-source models on multiple international competitions and benchmarks. It performs excellently in natural language reasoning, scoring 92.6 on the AIME 2025 test, close to GPT-5. On the IMO 2025 test, it solved Problem 3 in one attempt and gave partially correct answers on other problems, demonstrating high-order reasoning capabilities.
Ring-1T's main features
-
Powerful natural language reasoning capability: Scored 92.6 on the AIME 2025 test, close to GPT-5's 94.6, demonstrating powerful mathematical reasoning capability.
-
Efficient problem-solving capability: On the IMO 2025 test, solved Problem 3 in one attempt and gave partially correct answers on other problems, possessing high-order reasoning capabilities.
-
Competitiveness across multiple domains: Performs outstandingly on tasks such as HMMT 2025, LiveCodeBench v6, CodeForces, and ARC-AGI-1, demonstrating broad applicability.
-
Open-source collaboration: Code and weights are fully open source, released on the Hugging Face platform, facilitating community exploration and feedback and accelerating model iteration and improvement.
Ring-1T's technical principles
-
Architecture design: Adopts the Ling 2.0 MoE architecture, combined with a trillion-parameter scale, providing the model with powerful expressive capability and efficient computational efficiency.
-
Pretraining corpus: Pretraining was completed on 20T of high-quality corpus, ensuring the model can learn rich language knowledge and patterns.
-
Reinforcement learning training: Conducts RLVR training targeted at reasoning capabilities through the self-developed efficient reinforcement learning system ASystem, improving the model's reasoning and decision-making capabilities.
-
Continuous iteration: The model is still continuously training, constantly optimizing performance and solving current issues such as language mixing and reasoning repetition.
Ring-1T's project address
- Hugging Face model library：https://huggingface.co/inclusionAI/Ring-1T-preview
- Ling Chat：https://ling.tbox.cn/chat
Ring-1T's performance
- Mathematical reasoning: On the IMO 2025 test, Ring-1T solved Problems 1, 3, 4, and 5 in one attempt (silver medal level), and on the third attempt gave a near-full-score proof process for the geometry proof in Problem 2.
- Code generation: Scored 2092 on the Codeforces test, surpassing GPT-5 (High)'s 2073; ranks first among open-source models on programming benchmarks such as LiveCodeBench.
- Comprehensive capability: Leads in performance on tasks such as medical question answering (HealthBench) and creative writing (Creative Writing v3), with an 81.59% win rate on Arena-Hard-v2, approaching GPT-5's 82.91%.
Ring-1T's application scenarios
-
Natural language reasoning tasks: Scored 92.6 on the AIME 25 test, close to GPT-5's 94.6, demonstrating powerful mathematical reasoning capability.
-
Code generation and optimization: Surpassed GPT-5 on the CodeForces test with a score of 94.69, proving its outstanding code generation capability.
-
Multi-agent framework applications: Integrated with the multi-agent framework AWorld, it can be used to test and explore complex reasoning tasks.
-
Academic research and development: As the world's first open-source trillion-parameter reasoning large model, it provides researchers and developers with a high-performance, reproducible reasoning foundation, promoting transparency and collaborative innovation in the large model ecosystem.