Tencent Unveils T1 Inference Model, Outperforming DeepSeek

According to Tencent's announcement on its official WeChat account, the upgraded T1 model offers faster response speeds and superior handling of long text documents.

Tencent unveiled its T1 inference model, which outperforms DeepSeek on several metrics of knowledge and reasoning ability.

The post said T1 can "keep content logic clear and text neat and clean," and that the rate of "hallucination" — a phenomenon where AI generates incorrect information — is "extremely low."

The launch comes amid fierce competition in China's AI market, especially after DeepSeek introduced models with performance comparable to or superior to Western systems, at significantly lower costs.

Tencent previously released a beta version of T1 across platforms, including its AI assistant app Yuanbao.

The official version will be powered by Tencent's Turbo S platform language model, which was announced late last month. The company claims Turbo S processes queries faster than rival DeepSeek's R1 model.

A chart published in the post comparing the T1 model with DeepSeek R1 shows that Tencent's product excels on several metrics of knowledge and reasoning ability.

Tencent has been ramping up its investment in AI in recent months. On Thursday, the company announced plans to increase capital spending in 2025, after spending heavily on AI through 2024.

The T1 has scored high on several important artificial intelligence benchmarks, placing it among the top models worldwide.

Recent evaluations show that T1 scored 87.2 on MMLU-Pro, a dataset that measures basic reasoning in large language models. This score is second only to OpenAI o1.

Furthermore, T1 performed well on public benchmarks such as CEval, AIME, and Zebra Logic, which measure general knowledge and reasoning in both Chinese and English.

Source: https://www.baogiaothong.vn/tencent-trinh-lang-mo-hinh-suy-luan-t1-vuot-troi-so-voi-deepseek-192250322131611869.htm