Chinese language tech firm Tencent simply launched its newest giant language mannequin, Hunyuan Turbo S, that includes considerably sooner response instances with out sacrificing efficiency on advanced reasoning duties.
Tencent claims that its new AI doubles phrase era pace and cuts first-word delay by 44% in comparison with earlier fashions, in response to official info that the Chinese language tech large shared on Weibo.
The mannequin makes use of what seems to be a hybrid structure combining Mamba and Transformer applied sciences—the primary profitable integration of those approaches in a super-large Combination of Specialists (MoE) mannequin.
This technical fusion goals to unravel elementary issues which have plagued AI improvement: Mamba handles lengthy sequences effectively whereas Transformer captures advanced contexts, doubtlessly reducing each coaching and inference prices. Being hybrid signifies that the mannequin combines reasoning capabilities with the normal method of regular LLMs that present rapid response.
“The mixture and complement of quick pondering and sluggish pondering could make giant fashions clear up issues extra intelligently and effectively,” Tencent wrote when asserting the mannequin on its official WeChat channel. The corporate drew inspiration from human cognitive processes, designing Hunyuan Turbo S to supply immediate responses like human instinct whereas sustaining the analytical reasoning capabilities wanted for advanced issues.
Efficiency benchmarks present Hunyuan Turbo S matching or exceeding top-tier fashions throughout varied checks. It scored 89.5 on MMLU, barely above OpenAI’s GPT-4o, and achieved high scores in mathematical reasoning benchmarks MATH and AIME2024. For Chinese language language duties, it reached 70.8 on Chinese language-SimpleQA, outperforming DeepSeek’s 68.0. Nevertheless, it lagged in some areas like SimpleQA and LiveCodeBench, the place GPT-4o and Claude 3.5 carried out higher.
The discharge intensifies the continued AI competitors between Chinese language and American tech companies. DeepSeek, a Chinese language startup that has gained consideration for its cost-effective, high-performing fashions, has been placing strain on each Chinese language tech giants and American corporations like OpenAI with its extremely succesful and extremely environment friendly fashions.
DeepSeek’s fashions reportedly value round $6 million to coach and are extraordinarily low cost to run, charging round $1.10 per million tokens of output vs OpenAI’s GPT-4.5 and its wildly costly $150 per million output tokens.
Tencent priced Hunyuan Turbo S competitively at 0.8 yuan (roughly $0.11) per million tokens for enter and a pair of yuan ($0.28) per million tokens for output—considerably cheaper than earlier Turbo fashions. The mannequin is technically out there by way of API on Tencent Cloud, with the corporate providing a free one-week trial, however it’s nonetheless not out there for public obtain.
Regardless of the announcement, Hunyuan Turbo S is not but broadly accessible for obtain, however may be accessed by way of the Tencent Ingot Expertise web site. builders and companies want to hitch a ready listing by means of Tencent Cloud to realize entry to the mannequin’s API. The corporate hasn’t supplied a timeline for basic availability by way of Github.
The mannequin’s concentrate on pace might make it perfect for real-time functions like digital assistants and customer support bots—areas which can be very talked-about in China and by which Hunyuan Turbo S might provide vital benefits if it delivers on its promised capabilities.
Chinese language competitors within the AI area continues to warmth up, with the federal government pushing for extra adoption of native fashions. Past Tencent, Alibaba lately launched its newest state-of-the-art mannequin Qwen 2.5 Max, and startups like DeepSeek have launched more and more succesful fashions in latest months.
Edited by Andrew Hayward
Typically Clever E-newsletter
A weekly AI journey narrated by Gen, a generative AI mannequin.