Tencent unveils T1 reasoning model as AI race heats up in China

Tencent Holdings has introduced a new artificial intelligence (AI) reasoning model, Hunyuan T1, designed to compete with DeepSeek’s R1 in both performance and affordability.

Unveiled on Friday, T1 leverages large-scale reinforcement learning, the same technique DeepSeek used in its R1 model launched in January. 

T1 demonstrates strong performance across benchmarks

The official release follows a beta test of the T1 preview on Tencent’s chatbot, Yuanbao. The model scored 87.2 points on the Massive Multitask Language Understanding (MMLU) Pro benchmark, surpassing DeepSeek-R1’s 84 points but falling short of OpenAI’s o1, which achieved 89.3.

T1 also demonstrated strong results in other assessments, scoring 78.2 in the 2024 American Invitational Mathematics Examination (AIME), compared to R1’s 79.8 and o1’s 79.2. According to Tencent, it excelled in Chinese language proficiency, scoring 91.8 points in the C-Eval suite, matching R1 and outperforming o1’s 87.8.

The Tencent model also rivals DeepSeek in pricing, which is a primary advantage for the popular Chinese start-up. T1 charges 1 yuan ($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.

According to Tencent, it is the first in the industry to adopt a hybrid architecture, including Google’s Transformer, along with a generator-provided system by Carnegie Mellon University and Princeton University called Mamba. This significantly reduces training and inference costs and optimizes memory use during training compared to a fully Transformer-based architecture, the company states.

The company promotes T1 as “significantly reducing resource consumption while ensuring the ability to capture long text information,” allowing for 200% faster decoding Speeds.

A former JD.com big-data expert wrote the tech blog NCJRYDS and tested T1 and R1 on the same tasks. The individual then asked other large language models, including Anthropic’s Claude and OpenAI’s ChatGPT, to judge the results.

Tencent’s model lost the duel in writing an ancient Chinese poem but beat DeepSeek in interpreting a Chinese word in different contexts.

Another blog, GoPlayAI, gave the Tencent model four mathematical questions, and after spending five minutes on it, it got the hardest one wrong.

Tencent expands its AI strategy with a dual-core approach

Tencent, which operates China’s biggest social media app, WeChat, and the world’s largest video gaming business by revenue, is positioning AI as a new core revenue stream. The firm has integrated DeepSeek-R1 into its cloud platform and Yuanbao chatbot, offering an alternative alongside its own Hunyuan models.

Tencent chairman and CEO Pony Ma Huateng recently noted that he had great respect for DeepSeek for making “an independent, truly open-source and free product”.

Ma said Tencent has adopted a “double-core” strategy on AI that uses both DeepSeek and its own Yuanbao models. This approach is similar to how it has dominated the video gaming industry by promoting self-developed titles and those from independent studios.

Cryptopolitan Academy: Tired of market swings? Learn how DeFi can help you build steady passive income. Register Now