Ring-1T: Trillion-Parameter Model Trained with RLVR and RLHF (ant-ling.medium.com)

Ant Group’s inclusionAI team has open-sourced Ring-1T, a “thinking” trillion-parameter Mixture-of-Experts (MoE) model built on the Ling 2.0 architecture (Ling-1T-base). Ring-1T uses reinforcement learning with verifiable rewards (RLVR) plus RLHF to tune reasoning and general utility; it exposes 1T total parameters with ~50B active experts and supports up to a 128K context window. The release also documents system-level advances: Icepop (aka Popsicle) — a masked bi-directional truncation method — reduces train/inference distribution drift in long-cycle RL for MoE, while ASystem (SingleController + SPMD) and a unified memory pool enable zero-redundancy, low-latency weight exchange and large-scale RL training. The team has open-sourced the AReaL framework and shared demos of visualization and code-generation abilities. Ring-1T claims open-source state-of-the-art on hard reasoning benchmarks (AIME, HMMT, LiveCodeBench, ARC-AGI-1) and strong performance on Arena-Hard, HealthBench, and creative tasks. In live stress tests it reached IMO “silver” standard (solving 4 problems in one pass, near-perfect on another) and solved five problems in ICPC World Finals comparisons — competitive with closed models like GPT-5-thinking and Gemini-2.5-pro. The release is significant because it demonstrates RL-stabilized scaling of MoE to trillion parameters and provides infrastructure for community RL research, but the team notes remaining issues (benchmark contamination risks, identity/language mixing, repetition, and GQA attention inefficiencies) and ongoing optimization.