Ant releases the first open-source trillion-parameter inference model, Ring-1T (huggingface.co)

The Ling team at inclusionAI has open-sourced Ring-1T-preview, a trillion-parameter "thinking" inference model built on the efficient Mixture-of-Experts (MoE) Ling 2.0 architecture. Ring-1T was pre-trained on ~20T tokens and then post-trained with a reinforcement-learning-for-reasoning (RLVR) pipeline inside the team's ASystem (AReaL) framework, leveraging the disclosed "icepop" method. The preview weights are available on Hugging Face (inclusionAI/Ring-1T-preview) under an MIT license and are runnable with common Transformers workflows—supporting long generations (examples show max_new_tokens up to 8192). The team notes continued training and known issues such as language mixing, repetitive chains, and identity misperception. Why it matters: Ring-1T is one of the first openly available trillion-parameter models focused on robust natural-language reasoning, lowering the barrier for academic and developer experimentation with very-large-scale “thinking” models. Early evaluations are strong: a 92.6 score on AIME 2025 (approaching GPT-5-with-thinking at 94.6), competitive results on HMMT, code benchmarks (LiveCodeBench v6, CodeForces), ARC-AGI-1, and multi-agent IMO experiments using AWorld where Ring-1T solved a problem in one attempt and produced partial solutions on others. Open-sourcing this preview invites the community to probe reasoning limits, reproduce multi-agent and RL-based training methods, and help iterate on remaining failure modes.