Inside Ring-1T: Ant Engineers Solve Reinforcement Learning Bottlenecks at Trillion Scale

Why It Matters

AI leaders for technological and geopolitical advantage.

Summary

Ant Group unveiled Ring-1T, which it calls the first open-source reasoning model with one trillion parameters, optimized for math, logic, code generation and scientific problems and supporting up to 128,000 tokens. To train at trillion-parameter scale the company developed three new reinforcement-learning and scaling techniques—IcePop, C3PO++ and ASystem—that stabilize MoE training, parallelize rollout processing and enable asynchronous SPMD control, improving GPU utilization and training stability. In benchmarks Ring-1T ranked second to OpenAI’s GPT-5 and led all open-weight models (93.4% on the AIME 25 leaderboard and strong coding performance), underscoring China’s accelerating investment in large models and intensifying competition with U.S. AI leaders for technological and geopolitical advantage.