Ring
model paperAnt Group's reasoning-oriented (thinking) model family, built on the Ling backbone with reinforcement learning. The world's first open-source trillion-parameter thinking model.
Outputs 4
Ring-1T-preview
modelWorld's first open-source trillion-parameter thinking model. Preview release of the Ring reasoning series.
Architecture MOE
Parameters 1T
Ring-1T: Every Step Evolves
paperScaling reinforcement learning for trillion-scale thinking model.
Ring-2.5-1T
modelHybrid linear-architecture reasoning model that matches gold-medal standards in IMO/CMO 2025 competitions.
Architecture MOE
Parameters 1T
Ring-2.6-1T
modelTrillion-param MoE reasoning model (~63B active) with adaptive <strong>high</strong> / <strong>xhigh</strong> reasoning effort. 128K → 256K context (YaRN). Introduces <strong>Async RL + IcePop</strong> training: asynchronous policy-sampling/parameter-update decoupling for stable trillion-scale RL. Self-reported AIME26 95.83, PinchBench 87.60 (claims it beats GPT-5.4 xHigh and Gemini-3.1-Pro high), GPQA Diamond 88.27, ARC-AGI-V2 66.18. MIT licensed.
Architecture MOE
Parameters 1T
Active params 63B
Context window 262,144
License MIT