To compete with Chatgpt and Deepseek, the Chinese giant Tencent also launches an AI which “reason”
While we thought we were witnessing at the beginning of an AI battle between China and the United States, it is rather a real civil war that seems to be looming on the side of the former Middle Kingdom. Indeed, the more a week passes without a large group in the country presents one or more new models of language. To speak only of the most important, after the tidal wave Deepseekwe were suddenly Alibaba and his WanThen Baidu and his Erniewithout forgetting Manus. And it's not over, since it's now Tencent's turn, another Chinese giant, to enter the dance.
Clearly assertive ambitions with the T1
The group,who had announced his intentions from 2023has just presented the final version of its T1 reasoning model. This new model is faster and more efficient in processing complex and bulky textual documents. The launch was relayed through the official WeChat de Tencent account. According to the company, this model is capable of maintaining a logic of coherent content on long sequences and generating a “clean and neat” text. It would also display a very low hallucination rate. This T1 is also based on the Turbo S base, a language architecture unveiled by Tencent at the end of February. The group claims that this structure gives its new model a speed of execution greater than that of the DEEPSEEK R1 model, one of its main competitors in China. A data that Tencent did not fail to highlight, in particular to seduce companies and developers who have fallen under the thumb of Deepseek and its fairly irresistible performance/price ratio.
Convincing benchmarks (as always)
As usual, Tencent has published a comparative table in which he confronts the results of the new T1 with those of other competing models. According to the brand data, its T1 has obtained a score of 87.2 on the Benchmark MMLU-Pro, against 84 for the DEEPSEEK R1. The O1 model of Openai remains in front with a score of 89.3. But it is much more expensive than its Chinese competitors.
On the American mathematical test likes 2024, the T1 reached 78.2, a score slightly less than those of the R1 (79.8) and the O1 (79.2). On the other hand, on the Chinese benchmark C-Eval, the Tencent model is equal to the R1, both at 91.8 points, in front of the Openai model, which displays a score of 87.8. Obviously, these benchmarks are as always to be taken with tweezers.
Transform-Mamba, the hybrid approach
On the architectural level, Tencent engineers opted for a hybrid configuration called Transformer-Mamba. This combination combines the classic transforming structure with the Mamba system, more recent. According to the company, this coupling would significantly reduce memory consumption. Such an approach would notably allow better management of extensive contexts, without requiring an excessive material infrastructure. Tencent evokes a gain of 200 % in decoding speed compared to conventional architectures.
Another element put forward by Tencent: the learning strategy. The T1 model would be drawn to 96.7 % via a reinforcement method, a fairly atypical choice in the universe of large language models, where supervised learning remains the majority. The firm sees it as a lever for optimizing the accuracy and stability of the responses. Deepseek had also mainly proceeded in this way for its R1 model.
“DEEPSEEK” prices
To counter the latter's offensive on the tariffs, Tencent adopts a similar, even more aggressive policy. The cost of using the T1 is thus set at 1 yuan (approximately 0.13 euros) per million tokens as a starter, and 4 yuan (approximately 0.52 euros) only for the exit. For comparison, Deepseek invoices 1 yuan per million tokens during the day, 0.25 yuan at night for the entrance, and up to 16 yuans for the day release (4 yuan per night). Tencent clearly seems to try to capture part of the users seduced by the competitiveness of Deepseek, but who would like to enjoy the sweetest rates during the day too.
Available in the Yuanbao application
If the two reasoning models mentioned at the moment are undoubtedly competing, Tencent did not choose to replace its partners with its own tools. The T1 model has been integrated into its IA Yuanbao assistant, but the application also continues to offer the DEEPSEEK R1 model. A so -called “double” approach, assumed by Pony MA, the group's CEO, who publicly praised the open source orientation of Deepseek.
This approach will allow users to alternate between the two models according to their preferences or the specifics of their requests. Companies and developers using Tencent Cloud services also benefit from this flexibility: depending on the needs, it becomes possible to favor a model or to combine both, depending on the use cases.
Finally, note that Tencent does not come out of nowhere in terms of artificial intelligence. In 2024, the company indeed devoted $ 11.7 billion to its investments in artificial intelligence, against 3.4 billion the previous year. On Thursday, March 20, the day before the T1 announcement, the group said they wanted to further strengthen its investments in 2025, continuing the dynamics initiated last year.
