1AI fromTencent HunyuanWeChat public was informed that on March 21, Tencent mixed yuan officially launched the official version of the self-developed deep thinking model mixed yuan T1.
According to the official"T1" is fast in spitting out words, can reply in a second, and is also good at processing ultra-long texts, which is a strong self-developed product of Tencent.inference model. Through massive reinforcement learning, combined with specialized optimization of science challenges such as math, logical reasoning, science, and code, the official version of Hybrid T1 further enhances reasoning skills.
On common benchmarks that exemplify the foundational capabilities of inference models, theFor example, in the large language model evaluation enhancement dataset MMLU-PRO, the hybrid T1 scored 87.2 points, second only to o1The T1 has also been recognized as the industry's leading mathematical and logical reasoning model. In CEval, AIME, Zebra Logic, and other public benchmark tests of Chinese and English knowledge and competition-level mathematical and logical reasoning, Hybrid T1 also scored at the level of the industry's leading reasoning models.
"T1" also demonstrated very strong adaptability in multiple alignment tasks, command-following tasks, and tool utilization tasks.
Officially, the official version of the Hybrid T1 follows the innovative architecture of the Hybrid Turbo S and adopts the Hybrid-Mamba-Transformer fusion model. This is the first time in the industry that the Hybrid-Mamba architecture has been losslessly applied to a very large inference model. This architecture effectively reduces the computational complexity of the traditional Transformer structure and reduces the memory consumption of the KV-Cache, thus significantly reducing the training and inference costs.
Officials also claimed thatThe Hybrid T1 also demonstrates unique strengths in the area of ultra-long textual reasoning.The hybrid T1 can effectively solve the problem of context loss and long distance information dependency. Based on the excellent long text capture capability, Hybrid Mamba T1 can effectively solve the common context loss and long distance information dependency problems in long text reasoning. At the same time, the hybrid Mamba architecture is optimized for long sequence processing, which ensures the ability to capture long text information while significantly reducing resource consumption through efficient computation, and achieves a 2X increase in decoding speed with a similar number of activation parameters.
Tencent Hybrid T1 is now live: https://llm.hunyuan.tencent.com/#/chat/ hy-t1
In terms of API usage, the hybrid T1 is already online on the Tencent Cloud website, with an input price of $1 per million tokens and an output price of $4 per million tokens.
