Tencent has unveiled the Hunyuan T1 logical reasoning artificial intelligence model, developed using Large-scale Reinforcement Learning similar to DeepSeek’s R1 model. They have designed a hybrid model architecture using Google’s Transformer and Carnegie Mellon University’s Mamba, reducing training costs significantly.
Testing of the T1 model yielded an impressive MMLU score of 87.2, outperforming DeepSeek’s R1 at 84 but falling short of OpenAI’s o1. However, Tencent claims that the operational costs of T1 are lower than R1, priced at 1 million tokens for input and 4 million tokens for output, whereas R1’s pricing varies with time – 1 token for input and 16 tokens for output during daytime, dropping to 0.25 and 4 tokens during nighttime.
Hunyuan T1 is now available through Huggingface and Github, allowing users to experiment with it via the Hunyuan chatbot.
Source: Tencent via South China Morning Post
TLDR: Tencent introduces the Hunyuan T1 AI model, showcasing superior performance and cost-efficiency compared to existing models, offering access through Huggingface and Github.
Leave a Comment