1
NVIDIA Releases Rubin CPX GPU
NVIDIA released the Rubin CPX GPU in September 2025, designed specifically for large-scale contextual AI inference. This chip features an innovative architecture that divides AI inference into two stages: context processing and output generation, equipped with 128GB GDDR7 memory, providing 30 petaFLOPs of computing power, supporting long context processing of over 1 million tokens, achieving performance 6.5 times that of current flagship models. It is expected to be available by the end of 2026, primarily targeting scenarios such as code generation and video generation.
2
Tencent Open Sources Hunyuan Image Model 2.1
Tencent open-sourced its Hunyuan Image Model 2.1 on September 9, 2025. This model supports native 2K resolution image generation and complex prompt inputs of up to 1000 tokens, excelling in text semantic understanding and image-text embedding, accurately generating images with text in both Chinese and English. The concurrently open-sourced PromptEnhancer text rewriting model optimizes user instructions, enhancing the quality and expressiveness of generated images. The model is now available on platforms like Hugging Face and GitHub.
3
MagicLab Releases Full-Size Bipedal Humanoid Robot MagicBot Gen1
MagicLab has released the world’s first full-size general-purpose bipedal humanoid robot MagicBot Gen1 “Wheat”. Standing at 1.74 meters tall, it features 42 degrees of freedom and 350N.m joint torque, utilizing a three-layer intelligent architecture to achieve millisecond-level response, with a single charge lasting up to 5 hours. This robot integrates a large language model and a visual recognition system, supporting multilingual explanations, dynamic obstacle avoidance, and VIP welcoming, already applied in smart guiding scenarios in museums and science centers.
4
Bilibili Open Sources Self-Developed Speech Generation Model IndexTTS-2.0
Bilibili open-sourced its self-developed speech generation model IndexTTS-2.0 on September 8, 2025. This model adopts an autoregressive architecture and is the world’s first zero-shot text-to-speech (TTS) system that supports precise duration control, allowing users to adjust duration in milliseconds by specifying the number of tokens, completely solving the audio-visual synchronization problem. Its core breakthrough lies in the decoupling technology for emotion and timbre, allowing users to independently control the speaker’s timbre and emotional characteristics, and supporting intuitive emotional direction control through natural language commands. In multiple benchmark tests, its word error rate, speaker similarity, and emotional fidelity surpass current state-of-the-art models. The model code, weights, and technical report have been made public.
5
Shengshu Technology Launches Vidu Video Large Model Q1 Reference Image Function
Shengshu Technology officially launched the Vidu video large model Q1 reference image function on September 9, 2025. This function supports simultaneous input of 7 reference images, breaking through the technical challenges of multi-subject consistency, surpassing international models like Google’s Nano Banana in key metrics such as image consistency and fidelity. It supports 1080P HD output and enhanced Chinese semantic understanding, aiming to empower visual production needs in industries such as e-commerce and advertising.
6
Tencent Releases AI CLI Tool CodeBuddy Code
Tencent released the AI CLI tool CodeBuddy Code on September 9, 2025, allowing developers to drive the entire development process through natural language commands, achieving automated code generation, testing, and deployment. This tool does not require an invitation code and is open to all users, with the domestic version using the DeepSeek model for free, while the international version supports mainstream models like GPT and Gemini. It deeply integrates with the command line environment, significantly improving batch task and operation efficiency, aiding in enhancing R&D productivity.
7
Baidu Releases Wenxin Large Model X1.1 Deep Thinking Model
Baidu released the Wenxin Large Model X1.1 Deep Thinking Model on September 9, 2025. This model is based on the Wenxin Large Model 4.5 optimization and adopts an iterative mixed reinforcement learning training framework, achieving 34.8% improvement in factual accuracy, 12.5% increase in instruction adherence, and 9.6% enhancement in agent capabilities. The model is now available on Baidu Smart Cloud’s Qianfan platform and is open to enterprises and developers.
Previous Recommendations
Apple is secretly brewing a big move: self-developed AI search tool WKA! The Chinese Academy of Sciences open-sources “Shunxi 1.0”, and Fudan uses AI to crack the secrets of oracle bones.
OpenAI, SenseTime, and Atlassian all made big moves on the same day, while the Tsinghua team quietly dropped a deep-water bomb.
Every time you chat with ChatGPT, a city is “burning electricity” behind the scenes! In the next decade, whoever controls energy will control the AI world.
Please like
Please share
Please like
