Tencent

Tencent Hunyuan Team

In this paper, we introduce Hunyuan-Large, which is currently the largest open-source Transformer-based mixture of experts model, with a total of 389 billion parameters and 52 billion activation parameters, capable of handling up to 256K tokens. We conduct a thorough evaluation of Hunyuan-Large's superior performance across various benchmarks including language understanding and generation, logical reasoning, mathematical problem-solving, coding, long-context, and aggregated tasks, where it outperforms LLama3.1-70B and exhibits comparable performance when compared to the significantly larger LLama3.1-405B model. Key practice of Hunyuan-Large include large-scale synthetic data that is orders larger than in previous literature, a mixed expert routing strategy, a key-value cache compression technique, and an expert-specific learning rate strategy. Additionally, we also investigate the scaling laws and learning rate schedule of mixture of experts models, providing valuable insights and guidances for future model development and optimization. The code and checkpoints of Hunyuan-Large are released to facilitate future innovations and applications. Codes: https://github.com/Tencent/Hunyuan-Large Models: https://huggingface.co/tencent/Tencent-Hunyuan-Large

Fang Wang, Shaoying Li, Lin Liu, Feng Gao, Yanfen Feng, Zilong Chen

De-Zhi Ye, Jie Liu, Junwei Hu, Jia-Bin Fan, Bowen Tian, Hai-Jin Liang, Jin Ma

Relevance plays a crucial role in commercial search engines by identifying documents related to user queries and fulfilling their search needs. Traditional approaches employ encoder-only models like BERT, which process concatenated query-document pairs to predict relevance scores. While autoregressive large language models (LLMs) have revolutionized numerous NLP domains, their direct application to web-scale search systems presents significant challenges. On one hand, the relevance modeling capabilities of LLMs have not been fully explored. On the other, the high computational costs and inference times make deploying LLMs in online search systems, which demand extremely low latency, nearly impossible. In this work, we address these challenges through two key contributions. First, we develop a comprehensive evaluation framework to systematically assess the effectiveness of LLMs in query-document relevance ranking. By conducting assessment experiments to LLMs in four perspectives: ranking objectives, model size, domain-specific continuous pre-training, and the integration of prior knowledge, we identify the best resource allocation strategy given a restricted budget and develop practical LLMs in a more efficient way. Second, we propose a novel framework to transfer the capabilities of LLMs in the ranking aspect to existing BERT models to avoid directly deploying LLMs. Finally, to fully leverage the improvements in relevance ranking brought by LLMs, we successfully nearline deploy LLMs in Tencent QQ Browser search engine using query-based on-demand computing and quantization. Experiments on real-world datasets and online A/B tests demonstrate that our approach significantly enhances search engine performance while maintaining practical operational efficiency. Our findings provide actionable insights for integrating LLMs into production search engines.

Lu-Zhou Li, Zhi'an Zhang

There have been global regulatory debates and efforts aimed at holding digital platforms accountable for the public interest. Utilizing mixed methods of documentary research and expert interviews, this article examines how the Chinese digital giant Tencent leverages its adaptive capabilities for corporate social responsibility (CSR) initiatives in response to increasingly stringent regulation in China. Specifically, it explores Tencent’s organizational restructuring, including the establishment of the sustainable social value (SSV) division, as well as the accompanying corporate practices and discursive strategies aimed at enhancing its legitimacy. As an exemplar of the “institutional chameleon,” Tencent has deftly navigated regulatory waters through organizational metamorphosis. However, while Tencent’s enhanced CSR initiatives have certainly yielded some social benefits, the SSV initiatives are increasingly emerging as a new business frontier for the company, leveraging its vast infrastructural capacities to capitalize on the very social fabric it purports to support. This underscores the complexities and contradictions inherent in its CSR practices and the significance of the intrinsic logic of digital platforms as a modern business.

Zi-Hao Chen, Jiazhi Jiang, Jiangang Liu, Chao Zhang, Yuqi Diao, Yang Li, Hanmei Luo, Peng Chen

Stream services have gained immense popularity for facilitating real-time data processing, powering numerous applications within Tencent. Most of these streaming jobs are deployed at scale on the cloud to achieve higher availability and stability. Due to the dynamic nature of input data rates, cloud-native streaming services within Tencent experience both periodical and distributional workload imbalances. Vertical scaling mechanism provided on the cloud platform aids in more fine-grained automatic resource adjustments for the workload imbalances. However, current vertical autoscaling solutions on cloud platform face challenges in achieving timely and accurate scaling predictions, managing service interruptions caused by resource scaling, and efficiently handling various autoscaling abnormalities. These issues can hinder the solutions from meeting the service level objectives (SLO) of streaming jobs in Tencent. In this paper, we introduce Oceanus, an end-to-end solution designed to facilitate precise and automated resource adjustment through vertical autoscaling for cloud-native streaming services, while ensuring that their SLO are consistently met. Oceanus incorporates proactive ML-enhanced forecasting technique that learns from historical load patterns to accurately determine optimal scaling times, cross-layer in-situ resource adjustment mechanism that enables resource adjustments without interrupting stream processing and robust resolution workflow to ensure rapid recovery from unexpected issues after resource adjustment. We evaluate Oceanus in our production environment, and the experiment results have further validated our system design.