Tencent Upgrades Self-Developed Foundation Model Tencent Hunyuan and Deploys it to Internal Products

Translation: Tencent has released its third-quarter financial report for 2023, which includes updates on the development of AI large-scale model-related businesses. Tencent has expanded the parameters of its advertising AI models, enhancing the accuracy of targeting and attribution, and has provided generative AI tools to advertisers. Additionally, Tencent has upgraded its self-developed foundational model, Tencent Hunyuan, and deployed it across its internal products, making it available to external enterprise clients as well. Through the enhancement of its Model-as-a-Service (MaaS) capabilities, Tencent Cloud has integrated large-scale models into multiple products. Tencent Hunyuan has been integrated into over 180 business areas, widely utilized in advertising, gaming, content, and other sectors. Tencent Cloud is also driving the application of large-scale models across various industries, providing support for digital transformation initiatives.

High Demand Causes Strain on Computing Power WorkBuddy Completes Emergency Upgrade of Tencent Hunyuan Hy3 Model

WorkBuddy and Tencent Hunyuan joint project announced that due to the popularity of the initial launch of the Hunyuan Hy3 large model, computing resources experienced peak congestion. Since 10:00 a.m. on July 8th, a queue phenomenon occurred. The project team has completed an emergency computing power expansion and scheduling to ensure stable service operations.

AI Audio Editing Enters a New Era: Tencent Hunyuan Collaborates with Leading Institutions to Release the MMAE Benchmark. Current Model Precision in Audio Editing is Less Than 5%

Tencent Hunyuan, in collaboration with Shanghai Jiao Tong University, Nanyang Technological University of Singapore, Tianjin University, Peking University, and Fudan University, has launched the first general instruction-driven audio editing benchmark dataset, MMAE. This benchmark addresses the current limitations in AI's ability to edit audio, filling a gap in the field of audio generation and providing an important evaluation standard for multi-task audio editing research.

Tsinghua University and Tencent Hunyuan Win the MLSys2026 MoE Inference Challenge with a 4.1x Speedup on NPU

The Tsinghua University Storage Lab and the Tencent Hunyuan AI Infra team won the global championship in the MLSys2026 MoE Model Inference Optimization Challenge. To address the inference bottlenecks of the trillion-parameter mixture-of-experts (MoE) architecture on heterogeneous NPUs, the joint team designed a full-chain optimization solution, including the E-Shard strategy, PSUM three-dimensional tensor batch reading, and GEMV path, significantly improving performance.

Tencent Open-Sources Multilingual Translation Tool Hy-MT2, Lightweight Version Only 440MB for Local Execution, Mini Program Now Available

Tencent Hunyuan recently open-sourced the multilingual translation model Hy-MT2 and launched the "Tencent Hy Translation" mini program. This model family includes three sizes, supporting mutual translation among 33 languages and five ethnic languages/dialects. The lightweight Hy-MT2-1.8B uses Tencent's self-developed AngelSlim 1.25-bit extreme quantization technology, optimized for mobile devices, balancing high quality with efficiency.

Tencent Hunyuan Launches Its First Industrial-Level 2Bit Edge Model: Achieving a Performance Turnaround with a 0.3B Parameter Size

Tencent Hunyuan releases the ultra-small model HY-1.8B-2Bit, which reduces the equivalent parameter count to 0.3B through an industrial-level 2Bit quantization scheme, with memory usage of approximately 600MB and a size smaller than some mobile applications. This technological breakthrough solves the problem of significant precision loss in low-bit quantization, providing a new approach for efficient deployment of large models on consumer-grade hardware.