The Chinese technology giant’s latest offering, launched on Friday, leverages large-scale reinforcement learning, a technique also employed by DeepSeek in its R1 reasoning model, which launched in January.
T1 has achieved high scores in other benchmarks as well, including 78.2 in the American Invitational Mathematics Examination (AIME) 2024, slightly behind R1’s 79.8 and o1’s 79.2. In terms of Chinese language capabilities, T1 excelled with 91.8 points in the C-Eval suite evaluation, the same score as R1 and better than o1’s 87.8, according to Tencent.
It also rivals DeepSeek on pricing, which has been a primary advantage for the popular Chinese start-up. T1 charges 1 yuan (US$0.14) per 1 million tokens of input, while output costs 4 yuan per million tokens. The input rate is in line with R1, which charges 1 yuan per million tokens during daytime hours, and just 0.25 yuan overnight. The output pricing is also comparable, given R1’s daytime rate of 16 yuan per million tokens, which drops to 4 yuan overnight.
Tencent claimed to be the first in the industry to adopt a hybrid architecture combining Google’s Transformer and Mamba, developed by Carnegie Mellon University and Princeton University. Compared with pure Transformer architecture, the hybrid approach “significantly reduces training and inference costs” by cutting memory usage, according to the Chinese tech giant.
