Search AI Products and News

Yanyu Technology does not develop underlying large models. With several AI creation applications, it has risen rapidly in three years and has now launched its IPO in Hong Kong, with a valuation as high as 3 billion USD, sparking discussions on the business value and competitive barriers of 'AI application layer intermediaries'.

The Chinese animated film 'Niu Lai' earned only 7,711 yuan at the box office in ten days, but it was popularized by A-share investors due to its name's homophonic association with the phrase "bull market is coming." It topped 72 Weibo hot search topics on August 15th, with a single-day box office of 717,000 yuan and nearly 4.19 million yuan the next day. The production company has a registered capital of only 100,000 yuan and is a decoration company. The main creators are only the mother and daughter, and the director is not a professional, handling the script, direction, editing, and voice acting alone. The film has been criticized for being crude and low-quality, with investors' pursuit of good fortune becoming a collective performance art.

Xiaohongshu released the open-source large model dots3-note preview, with a total of 280B parameters and 16B activated. It supports text, vision, and voice modalities. Its reasoning, Agent, and multimodal performance are close to larger models. On the same day, Huawei Ascend Atlas 800A3 and 900A3 SuperPoD completed efficient compatibility, and deployment was achieved through vLLM Ascend.

NVIDIA is discussing with SB Energy, a clean energy company supported by SoftBank, to invest up to 3 billion USD as a core part of its participation in the OpenAI data center transaction. The move aims to accelerate the construction of AI infrastructure and deeply support the AI industry chain with substantial financial power.

MiniMax launches MiniMax-Music3, generating up to 5-min songs from lyrics and music description, outputting 32kHz 16-bit stereo WAV. It uses a hierarchical autoregressive architecture: an 8B Global LLM predicts first RVQ codebook per frame for long-range structure, initialized from Qwen3-8B.....

Tencent reshuffles LLM team: Xu Can, senior Hunyuan researcher, moves to WeChat Group's WeLM team, focusing on post-training and agent development. He created WizardLM and proposed Evol-Instruct, with deep expertise in synthetic data and reinforcement learning. The move comes as WeChat AI accelerates deployment.....

CLUE's SuperCLUE-Terminal benchmark ranks DeepSeek-V4-Pro-0813 second with 51.52, behind Kimi K3, with very low cost and strong value. It targets Chinese developers, uses real local coding tasks, and runs models via Claude Code for fairness.....

Tencent's second-quarter financial report shows that AI has become the core focus: the Hunyuan large model is rapidly iterating, from Hy3 preview to the official version, and plans to launch Hy4 with even larger parameters; the office intelligent agent has initially established an advantage, with multiple products making progress; the Weixiao AI assistant for WeChat was launched in June and is now expanding gray testing. The fundamentals remain stable, with revenue reaching 204.79 billion yuan, showing growth year-on-year.

Tencent Q2: Hunyuan model Hy3 weekly calls surged 68x, Hy4 upcoming. AI office tool WorkBuddy had over 20M monthly visits, WeChat AI 'Xiaowei' in grey testing. Game revenue +11% YoY, local games +17%.....

Supercomputing Internet AI Community releases 'Light and Shadow Creativity', a free AI video agent. Powered by the full MiniMax-H3 model on domestic computing, it requires no local setup. Accepts text and image prompts, generating up to 15s 768P HD video per run.....

Last week, global AI large model calls reached 69 trillion tokens, up 21.48% WoW. China's weekly calls hit 34.25 trillion tokens, up 21.76%, maintaining a strong lead; the US had 9.17 trillion. China shows a clear advantage in global AI computing and applications.....

The Xiaomi Smart Camera 4Max AI Zoom Edition is officially on sale, priced at 739 yuan on JD.com. The core upgrade features the first AI care model from Xiaomi and a 3T four-core chip, tripling the computing power. It moves beyond traditional 'motion detection' alerts, supporting more detailed behavior recognition with the AI large model to improve monitoring accuracy.

ByteDance is discussing building China's largest model with over 5 trillion parameters, far exceeding existing competitors. The plan is still early, led by Xiang Liang and Shen Ke, and the Seed department is being restructured for this.....

MiniMax released open-source multimodal model H3, topping multiple authoritative benchmarks and the Hugging Face trending list, highlighting China's strong open-source AI capabilities. Within 24 hours, over 100 partners completed integration, demonstrating remarkable ecosystem momentum.....

New additions to the Hong Kong Stock Connect boost AI sector, with MiniMax shares rising over 10% after inclusion. The company also open-sourced a new generation of general multimodal generative model this week, advancing both capital and technology.....

Tencent Hunyuan launches Hy ASR3.0 Preview, built on the large language model Hy3. It integrates high-accuracy recognition with deep semantic understanding, enabling a shift from word-by-word transcription to contextual comprehension. Covering ten major dialect regions, it excels in long audio scenarios without relying on standard Mandarin, allowing AI to truly understand speech.....

Mistral AI released the open-source content moderation model Shieldstral, with 3B parameters, under the Apache 2.0 license, supporting 12 languages and running on a single 16GB GPU. It claims performance comparable to 7 times larger models and achieves SOTA in multimodal moderation, but points out that most protection models have a fixed harm classification system.

Tencent Hunyuan launches Hy ASR3.0preview, integrating Hy3 LLM's semantic understanding into speech recognition for context-aware, direct transcription. On open benchmarks, word error rates reach 3.34% for Mandarin, 2.62% for English, and 3.12% for Cantonese, achieving around 3% overall with impressive multi-dimensional performance.....

The Swiftlet project uses Swift + Metal to enable MoE models like Qwen3-Next to run with minimal size on Mac: only a small dense core is kept in memory, and expert weights are stored on SSD and loaded on-demand. On an M5 Mac, the 4-bit version of Qwen3.6-35B-A3B occupies only 18GB of disk space, with a peak memory of 2.6GB and a decoding speed of 7–11 tok/s.

CodeBuddy and WorkBuddy have extended the free access period for the Hy3 large language model to August 31, 2026, due to high user demand. Hy3 supports text-only tasks with no multimodal capabilities. If users attempt image or video generation via these products, the system will automatically switch to appropriate multimodal models and deduct credits as normal.....