Tencent Introduces New AI Model to Compete with DeepSeek-R1

The Chinese technology giant Tencent has launched a new AI model that aims to rival the performance of DeepSeek's R1 reasoning model. This new offering, introduced on Friday, utilizes large-scale reinforcement learning, a method that DeepSeek also implements in its model.

This release is the official version following a successful beta phase known as T1-preview on Tencent's chatbot Yuanbao. In benchmark tests, T1 achieved an impressive score of 87.2 points on the Massive Multitask Language Understanding (MMLU) Pro test, which evaluates a model's knowledge. This score exceeds DeepSeek-R1's benchmark of 84 points but still falls short of the 89.3 points recorded by OpenAI’s o1, a model released in December.

In other assessments, T1 showed strong capabilities, scoring 78.2 points in the American Invitational Mathematics Examination (AIME) 2024, which is slightly lower than R1’s score of 79.8, while OpenAI’s o1 managed 79.2 points. Tencent's T1 also excelled in evaluations of Chinese language prowess, scoring 91.8 points in the C-Eval suite, matching DeepSeek-R1 and surpassing o1, which earned 87.8 points.

On a pricing front, Tencent's T1 model competes effectively with DeepSeek. T1 charges 1 yuan (approximately US$0.14) for every 1 million tokens of input, while the output price is set at 4 yuan per million tokens. This pricing structure mirrors that of R1, which charges 1 yuan during the day and 0.25 yuan overnight for input, while its output costs 16 yuan per million tokens, decreasing to 4 yuan at night.

Tencent claims to be a pioneer in utilizing a hybrid architecture that combines Google's Transformer technology with Mamba, developed by Carnegie Mellon University and Princeton University. This innovative approach reportedly leads to a significant reduction in training and inference costs by minimizing memory usage compared to traditional pure Transformer models.

AI, Tencent, DeepSeek