Ring-1T

Ring-1T is a trillion-parameter reasoning model developed by InclusionAI. It is built on the Ling 2.0 architecture, employing a sparse Mixture-of-Experts (MoE) configuration that contains 1 trillion total parameters with approximately 50 billion parameters active per token. The model is designed for high-intensity reasoning tasks and supports a context window of 128,000 tokens.

The development of Ring-1T focused on scaling reinforcement learning (RL) for large-scale models. It was trained using Reinforcement Learning from Verifiable Rewards (RLVR) and Reinforcement Learning from Human Feedback (RLHF) to enhance its logical consistency and problem-solving abilities. To address training instability at the trillion-parameter scale, the creators implemented Icepop, a stabilization framework that uses token-level discrepancy masking to align training and inference behavior.

The model demonstrates strong performance in symbolic logic, programming, and advanced mathematics. In standardized evaluations, it has achieved results comparable to a silver medal on the 2025 International Mathematical Olympiad (IMO) and shown competitive proficiency in ICPC-level coding challenges. It is released as part of a series of thinking models that prioritize verifiable correctness over simple text generation.