AI快讯｜AI News

Trae v1.0.8 version is officially released, adding full support for Claude 3.7 Sonnet models, providing users with more powerful AI generation capabilities. Resources are limited, response may be delayed, the service is currently free, users can experience more efficient and intelligent AI features.

OpenAI launches GPT-4.5 as a research preview version of the most powerful chat model, currently available only to Pro users and developers.GPT-4.5 improves pattern recognition, knowledge breadth, and creativity by scaling up unsupervised learning, with fewer hallucinations and more natural conversations. It excels in writing, programming and problem solving with higher emotional intelligence and better understanding of human intent.

DeepSeek released Fire-Flyer File System (3FS) and Smallpond data processing framework on the last day of Open Source Week. 3FS is a high-performance parallel file system designed for AI training and inference, supporting high-throughput data access in large-scale clusters, with peak read throughput of 6.6 TiB/s. 3FS is built on 3FS and DuckDB to support high-efficiency data processing. Smallpond, built on 3FS and DuckDB, supports efficient data processing and scales to petabyte-scale datasets.

Baidu announced that Wenxin Big Model 4.5 will be officially launched on March 16th. On the same day two years ago, Baidu released the world’s first generative AI product, Wenshin Yiyin. The upgraded Wenxin Big Model 4.5 significantly improves on the basic modeling capabilities, adding new native multimodal and deep thinking capabilities.

Honor 100 announced that it has entered into a cooperation with Ali in the field of AI, and a number of language, visual understanding and image generation models such as Qwen and Wan have been connected to the Honor YOYO Intelligent Body. Users can experience the relevant AI functions through the Magic7 series of cell phones.YOYO assistant has been online DeepSeek-R1 full-blooded version, which supports the deep thinking mode, and the answer is more accurate and timely.

Tencent released a new generation of fast-thinking model, Hunyuan Turbo S. The model excels in knowledge, mathematics and science, and creativity, and realizes the function of “second reply”, doubles the speed of spitting out words, and reduces the latency of the first word by 44%. Turbo S adopts the Hybrid-Mamba-Transformer convergence architecture, effectively reducing computational complexity and deployment costs.

DeepSeek open source three parallel computing optimization techniques, DualPipe is a two-way pipeline parallel algorithms, through the complete overlap of computation and communication, significantly reduce the pipeline bubble, optimize the utilization of GPU resources. EPLB (Expert Parallel Load Balancer) through the redundancy of expert policy and load balancing techniques, to solve the problem of uneven load on GPUs in expert parallelism. Profile-data provides detailed performance analysis data to help developers gain a deeper understanding of parallel policy optimization.

Anthropic has announced a fully open GitHub integration for Claude AI for Free, Pro and Teams users. The integration allows users to sync their GitHub codebase to Claude, allowing it to understand code context and assist in testing, debugging and optimizing code.

Microsoft has open sourced two small models, Phi-4-Multimodal and Phi-4-Mini. Phi-4-Multimodal is a multimodal model that integrates textual, visual, and speech inputs, and utilizes “hybrid LoRA” technology for superior performance. Phi-4-Mini is a 3.8-billion-parameter language model that excels at math and coding tasks and outperforms models of its size.