Ant Group Open-Sources “Ring-1T-preview,” a Trillion-Parameter Reasoning Model
Want to read in a language you're more familiar with?
Ant Group open-sources Ring-1T-preview, a trillion-parameter reasoning model reporting SOTA benchmarks and strong code generation (per Ant).
On Sept. 30, 2025, Ant Group announced the open-source release of Ring-1T-preview, a trillion-parameter natural-language reasoning model and the preview build of its forthcoming Ring-1T system. Ant says the model delivers state-of-the-art results across several public benchmarks and, in some cases, approaches or exceeds leading closed models.
According to Ant’s Bailing team, Ring-1T-preview scores 92.6 on AIME-25, topping all known open models and Gemini 2.5 Pro, and landing just shy of GPT-5 (no tools) at 94.6. On Codeforces, the model reports 94.69, which Ant claims surpasses GPT-5. Ring-1T-preview also ranks first among open models on LiveCodeBench and ARC-AGI-v1, based on Ant’s disclosures.
The research team further trialed the model on IMO-25 math tasks: Ring-1T-preview was able to fully solve the third problem in one pass and derive partially correct answers for problems 1, 2, 4, and 5, Ant said. The company is continuing post-training on its Ling 2.0–family 1T base model to push reasoning performance while training the full Ring-1T release.
Ant describes Ring-1T-preview as optimized for reasoning and code generation rather than general chat, emphasizing long-horizon problem solving. Source code and model artifacts are being published openly, positioning Ring-1T-preview as the first open-source trillion-parameter reasoning model, per the announcement.
