Tencent Launches Hunyuan Turbo S: A Challenge in the AI Landscape

Chinese technology company Tencent has unveiled its latest large language model, known as Hunyuan Turbo S. This new model boasts significantly improved response times while still performing well on complex reasoning tasks.

Tencent reports that Hunyuan Turbo S doubles its word generation speed and reduces the first-word delay by 44% compared to earlier models, based on information shared on Weibo.

The model incorporates what seems to be a hybrid architecture, combining Mamba and Transformer technologies. This marks the first successful integration of these technologies in a super-large Mixture of Experts (MoE) model.

This innovative technical blend aims to address challenges that have historically hindered AI evolution. The Mamba technology efficiently manages long sequences, while the Transformer structure adeptly captures intricate contexts, which may notably reduce training and inference costs. By employing a hybrid framework, the model merges reasoning skills with the immediate response capabilities typical of standard large language models (LLMs).

"The combination of quick thinking and deliberative reasoning allows large models to address challenges with greater intelligence and efficiency," Tencent stated during the model's announcement on its official WeChat channel. The design of Hunyuan Turbo S was inspired by human cognitive processes, aiming to provide instantaneous responses similar to human intuition, while retaining the analytical reasoning needed for more sophisticated queries.

Performance evaluations indicate that Hunyuan Turbo S either meets or surpasses leading models in various assessments. It achieved a score of 89.5 on the MMLU benchmark, slightly edging out OpenAI's GPT-4o. The model also performed remarkably well in mathematical reasoning tests, excelling in MATH and AIME2024 benchmarks. In tests focused on the Chinese language, it scored 70.8 on Chinese-SimpleQA, outperforming DeepSeek's score of 68.0. However, it fell short in certain evaluations such as SimpleQA and LiveCodeBench, where GPT-4o and Claude 3.5 demonstrated superior performance.

The introduction of Hunyuan Turbo S further escalates the ongoing AI competition between Chinese and American technology firms. DeepSeek, a Chinese startup, has garnered attention for creating economical yet high-performing models, exerting pressure on established firms in both China and the U.S., including OpenAI.

Training DeepSeek's models reportedly costs around $6 million to produce, with operational costs that are impressively low—about $1.10 for every million tokens generated compared to OpenAI’s pricier GPT-4.5, which charges around $150 per million tokens.

Tencent has strategically priced Hunyuan Turbo S at 0.8 yuan (approximately $0.11) per million tokens for input and 2 yuan ($0.28) per million tokens for output, making it considerably cheaper than previous Turbo versions. Currently, the model is accessible via API on Tencent Cloud, where a free one-week trial is offered. However, it is not yet available for public download.

While Hunyuan Turbo S is not broadly downloadable, access can be obtained through the Tencent Ingot Experience site. Developers and businesses interested in using the model will need to join a waiting list via Tencent Cloud. The company has not announced a timeline for wider availability nor for inclusion on GitHub.

The model’s emphasis on rapid response times renders it highly suitable for real-time applications such as virtual assistants and customer service bots—sectors that are gaining immense popularity in China, where Hunyuan Turbo S could excel if it fulfills its touted capabilities.

The competition within the Chinese AI market is intensifying, driven by government incentives to promote local model adoption. Apart from Tencent, other companies like Alibaba have launched advanced models, such as Qwen 2.5 Max, while startups like DeepSeek continue to develop increasingly proficient models.

AI, Tencent, HunyuanTurboS