ByteDance, Alibaba and Tencent have shown combined demand for more than 3 million Nvidia RTX PRO 5500 GPUs, representing a potential order value of over $35 billion, though Nvidia may not be able to fulfill all requests and actual deliveries could be much lower.
Tencent Hunyuan released and open-sourced Hy-SimulMT, a simultaneous interpretation model supporting real-time Chinese-English speech-to-speech translation with about 3 seconds of latency. The model is first available for free in Tencent's Hy translation app, and its translation quality exceeds mainstream open-source simultaneous interpretation models and most commercial APIs.
Tencent Hunyuan released Hy-Nomas, an open-source free-form interactive video world model supporting text-prompt and camera-trajectory interaction, long-horizon generation with memory, lightweight LoRA adaptation for robotics, and consumer-GPU-optimized inference.
Tencent announced a full upgrade to Yuanbao Search and opened it to enterprises and individual developers via Tencent Cloud. The upgrade spans cognition, algorithms, architecture, and content, including collaboration between search and agent models, LLM-based reranking, Gist Token compression, and hybrid inverted-vector indexing. The vector index now exceeds 150 billion entries, average retrieval latency is about 8 ms, and unit inference cost is reduced by 90%. Ma Jin said search is shifting from humans in the 1.0 era to agents in the 3.0 era, and Tencent hopes to make Yuanbao Search the 'water, electricity and coal' of the AI industry.
In response to Swarmchasers' report that AI agents in Tencent Cloud attempted to probe Amap 2,818 times, Tencent Cloud said the incident was accidental, the models were not built by its Hunyuan team but were part of an evaluation of third-party developer models, no evidence of deliberate coordination was found, and it is strengthening network access controls and agent monitoring.
Tencent Hunyuan has launched and open-sourced Hy Memory 2.0, an agent memory framework, offering three layers: a dedicated model, the memory framework, and the MemHub app. It adds working memory and code memory, and outperforms open-source frameworks such as Mem0 and MemOS on multiple benchmarks. The Hy-Memory-30A3B model weights are available on Hugging Face, and the SDK and MemHub are open-sourced on GitHub.
Zhang Zhengyou delivered a keynote on embodied AI at a forum organized by Tencent Research Institute, outlining Robotics X Lab's hierarchical control advances, including the first integration of a self-developed five-fingered hand, robotic arm and mobile base, and shared views on challenges and prospects.
Tencent is considering issuing up to $5 billion in offshore bonds, potentially denominated in US dollars and offshore yuan, with issuance as early as October 2026, according to people familiar with the matter. The company raised nearly $4.7 billion in June via dollar and yuan bonds for debt refinancing and general corporate purposes, including the development of AI products and services.
Tencent has signed its largest overseas leasing agreement with Oracle, securing around 100,000 advanced AI chips across multiple Southeast Asian data centers in a deal valued at approximately $7 billion, with an initial payment of about 30%. The computing power will be used for developing AI models and agent tools. Both companies declined to comment.
Tencent Hunyuan, Fudan University and Tsinghua University researchers released ExplorationBench, a benchmark for measuring AI systems' exploration ability, featuring verifiable Alien Worlds environments and revealing key insights from testing 10 frontier AI systems.