TLDR: The Ling Team at Inclusion AI has introduced Ring-1T, the first open-source thinking model with a trillion parameters. It features 1 trillion total parameters and activates approximately 50 billion per token. The model addresses training challenges through three innovations: IcePop for stable RL training, C3PO++ for efficient long rollouts, and ASystem, a high-performance RL framework. Ring-1T achieves breakthrough results on benchmarks like AIME-2025 (93.4), HMMT-2025 (86.72), CodeForces (2088), and ARC-AGI-v1 (55.94), notably earning a silver medal-level result at IMO-2025. This release marks a significant step in democratizing large-scale reasoning AI.
The world of artificial intelligence has just witnessed a significant leap forward with the introduction of Ring-1T, a groundbreaking thinking model developed by the Ling Team at Inclusion AI. This model is the first open-source AI of its kind to boast a trillion-scale parameter count, setting a new benchmark for large-scale reasoning intelligence.
Ring-1T operates with a staggering 1 trillion total parameters, activating approximately 50 billion parameters per token. Training a model of this magnitude presents immense challenges, including maintaining stability during reinforcement learning (RL), efficiently processing long sequences of data, and overcoming systemic bottlenecks in the RL framework. To tackle these hurdles, the Ling Team pioneered three interconnected innovations.
Innovations Driving Trillion-Scale AI
The first innovation is **IcePop**, a method designed to stabilize RL training. It addresses the instability caused by mismatches between training and inference processes by using token-level discrepancy masking and clipping. This technique selectively corrects unstable contributions while ensuring efficient updates, leading to a more stable training environment.
Next is **C3PO++**, which significantly enhances resource utilization, especially for handling long sequences of data under a token budget. It achieves high time efficiency by dynamically partitioning these long rollouts, preventing computational resources from sitting idle and optimizing the reuse of data through batched continuation.
Finally, **ASystem** is a high-performance RL framework specifically engineered to overcome the systemic bottlenecks that typically impede the training of trillion-parameter models. Its architecture is built around a unified execution environment, featuring components like Hybrid Runtime for seamless training and inference, AMem for optimized GPU memory management, AState for high-speed weight synchronization, and ASandbox for rapid, isolated scenario validation. This robust framework ensures stability and efficiency for demanding workloads.
Also Read:
- Advancing Language Models in Solving Complex Optimization Problems
- Boosting LLM Reasoning and Clarity with Multi-Agent Debates
Breakthrough Performance Across Benchmarks
Ring-1T has delivered exceptional results across a variety of critical benchmarks, showcasing its advanced reasoning capabilities. It scored 93.4 on AIME-2025, 86.72 on HMMT-2025, and achieved a rating of 2088 on CodeForces. Furthermore, it reached 55.94 on ARC-AGI-v1. A particularly notable achievement is its silver medal-level performance at the IMO-2025, demonstrating its ability to solve complex mathematical problems through pure natural language reasoning, without relying on external tools or code generation.
The model’s training pipeline involves several stages: Long Chain-of-Thought Supervised Fine-Tuning (Long-CoT SFT) to build foundational reasoning, Reasoning RL using verifiable rewards (RLVR) for complex tasks, and General RL with human feedback (RLHF) to enhance human alignment and overall usability.
By making the complete 1T parameter Mixture-of-Experts (MoE) model open-source, the Ling Team aims to democratize access to cutting-edge reasoning intelligence, providing the research community with a new baseline for open-source model performance. For more detailed information, you can refer to the original research paper.
