Core Insights - Ant Group has launched its first trillion-parameter model, Ring-1T-preview, on Hugging Face, showcasing advanced natural language reasoning capabilities [3] - The model achieved a score of 92.6 in the AIME 25 test, surpassing all known open-source models and Gemini 2.5 Pro, and is close to GPT-5's score of 94.6 [2] - In the CodeForces test, Ring-1T-preview scored 94.69, exceeding GPT-5 [2] - The model ranks first among open-source models in authoritative lists such as LiveCodeBench and ARC-AGI-v1 [2] Technical Details - Ring-1T-preview utilizes the Ling-2.0 MoE architecture and has been pre-trained on 20 trillion tokens [2] - The model incorporates the previously released icepop method and has undergone RLVR training within Ant Group's self-developed efficient reinforcement learning system, ASystem [2] - The AReaL framework of ASystem has been open-sourced, indicating a commitment to transparency and collaboration in AI development [2]
蚂蚁推出开源万亿参数推理大模型 已上线Hugging Face
Huan Qiu Wang·2025-09-30 04:02