Search AI Products and News

DeepSeek affiliate company publicly disclosed a patent for deploying large language models, innovatively utilizing a distributed architecture: deploying the prefill and decode stages on high-performance computing and large-memory machines respectively. This method balances the load, reduces computational idle time, significantly lowering latency and increasing throughput. The patent highlights optimizations in system scalability and fault tolerance. Its MoE language model, DeepSeek-V3, has 671 billion parameters, with 37 billion parameters activated per token, which will promote the practical application of AI technology across industries. The core breakthrough lies in combining hardware resources effectively.

Tsinghua & partners open-sourced MOSS-TTSD, a bilingual speech model based on Qwen3-1.7B. Features XY-Tokenizer for 1kbps low-bitrate quality, zero-shot cloning, and 960s generation. Outperforms MoonCast in Chinese metrics.....

Black Forest Labs & Krea launch open-source image model FLUX.1-Krea, addressing AI's artificial look via 12B-parameter diffusion. Enhanced training improves realism in lighting/colors, compatible with FLUX ecosystem for commercial use.....

Amazon CEO Andy Jassy plans to integrate ads into AI assistant Alexa+, recommending products via dialogue. Despite user enthusiasm for Alexa+'s upgrade, AI ad chats are novel. Google and OpenAI are also exploring AI ads. Amazon has invested $31.4B in AI, focusing on chips and data centers, while tackling challenges like AI hallucinations and privacy.....

San Francisco AI startup Deep Cogito launches the Cogito v2 series of open-source large models, focusing on machine intuition and self-improvement capabilities. The series includes four dense and mixture-of-experts models with 70B to 671B parameters. The flagship 671B MoE model has a reasoning efficiency that is 60% higher than DeepSeek R1, and its performance is close to Claude4Opus. Its core technology involves internalizing reasoning paths into model weights through introspective reasoning, enabling continuous evolution while in use. Tests show that the model can

Poe launches a developer API supporting 100+ multimodal AI models, offering subscription plans ($4.99-$249.99/month) and pay-as-you-go credits ($30/1M). Developers can integrate AI into tools, with budget management features coming soon.....

The StepZen team announced that their latest generation foundational large model, Step3, is officially open-sourced. Step3 is a model designed for enterprises and developers who pursue the optimal balance between performance and cost, aiming to create the most suitable model for the inference era. The open-source address of this model includes GitHub, Hugging Face, and ModelScope, allowing developers to freely download and experience it. Step3 adopts an MoE architecture, with a total parameter count of 321 billion and an activated parameter count of 38 billion. It

The Qwen3-Coder series has welcomed a new member - Qwen3-Coder-Flash, a programming model affectionately called "a dessert-level" by developers. With its outstanding performance and efficient operation speed, this model brings new surprises to the field of programming. The full name of Qwen3-Coder-Flash is Qwen3-Coder-30B-A3B-Instruct. It performs excellently in terms of performance and efficiency, especially in agentic coding and browsing

The Aliyun Tongyi Lab open-sources the WebAgent intelligent agent project, which includes two core components: WebShaper and WebSailor. WebShaper addresses AI reasoning challenges using a formal-driven data synthesis method. The WebSailor-72B model surpasses most closed-source models in the BrowseComp evaluation. The project builds a complete ecosystem through the WebDancer training framework and WebWalker evaluation tool, supporting complex data crawling and analysis within 10 minutes. After open-sourcing,

The paper by Dr. Liang Wenfeng's team at DeepSeek in collaboration with Peking University has won the ACL 2025 Best Paper Award. This research proposes the Native Sparse Attention (NSA) mechanism, which achieves algorithm-hardware co-optimization through three parallel attention branches: compression, selection, and sliding. It improves the processing speed of long texts by 11 times and outperforms traditional full-attention models. NSA supports a context length of up to 1 million tokens, achieving an 11.6 times decoding speed improvement on 64k text, and a 27B parameter model demonstrates excellent performance on multiple benchmark tests.

Quest Mobile reports top 20 fastest-growing apps (June 2025). AIGC's DeepSeek leads with 163M MAU, Doubao hits 141M (+410% YoY). Hongguo (212M) and UnionPay (188M) lead short videos & payments. List reflects AIGC and online video growth potential.....

OWL team open-sourced Eigent, a multi-agent collaboration tool based on CAMEL and OWL frameworks. It enhances complex task efficiency via parallel sub-task processing, supports multiple LLMs and data integration, and promotes community development with modular design and standardized protocols.....

Kunlun open-sources Skywork UniPic, a 1.5B-parameter multimodal model integrating image understanding, generation, and editing. It unifies multimodal tasks with end-to-end pretraining, offering strong performance comparable to larger models.....

On July 30, Kunlun Infinite officially launched and open-sourced the autoregressive multi-modal unified pre-training model Skywork UniPic. The model integrates three core capabilities - image understanding, text-to-image generation, and image editing - in a single architecture. It demonstrates excellent generality and transferability through end-to-end pre-training on large-scale high-quality data.

Shenzhen Leju Robotics has secured a new round of funding, with its registered capital increased to 2.27 million yuan, and investors including Dongfang Jinggong. The company, established in 2016, has accumulated over tens of millions in financing. It has transitioned from educational robots to the industrial field. Its Aelos educational robot has been adopted by more than 4,000 schools. In 2023, it launched a humanoid robot based on open-source HarmonyOS, and collaborated with Huawei Cloud and China Mobile to develop 5G-A robots, enhancing application capabilities in industrial scenarios. The company plans to deliver 1,000 humanoid robots by 2025, addressing labor shortages in manufacturing and promoting intelligent production.

China Mobile launched the Jiutian foundational large model 3.0 at the 2025 World Artificial Intelligence Conference, achieving three major breakthroughs: 1) The Jiutian Zhongqing Language Large Model adopts a trillion-parameter MoE architecture, improving reasoning capabilities by 35%, and leading globally in evaluations such as GPQA; 2) The Jiutian Shanzhi Multimodal Large Model innovates spatiotemporal modeling technology, significantly enhancing image and video generation and understanding capabilities; 3) An open-source strategy is implemented, with specialized models and evaluation systems for mathematics, coding, and other areas released to promote industry ecosystem development. The model features high security and full domestication, demonstrating China's strength.

Shanghai AI Lab releases the open-source scientific multimodal large model Intern-S1, surpassing existing top models in scientific research fields such as chemistry and materials. The model has cross-modal scientific analysis capabilities, able to interpret complex data such as molecular formulas and protein structures, and also launches a dedicated scientific discovery platform called Intern-Discovery. Intern-S1 will promote the development of multi-agent disease research systems and continue to open-source to support scientific progress.

At the 2025 World Artificial Intelligence Conference, Wang Ai, Chief Product Officer of Honor, demonstrated the technical breakthroughs of the MagicGUI large model. This open-source large model enables the transformation of an intent-driven operating system, solves the challenges of cross-scenario adaptability, and builds an open AI terminal ecosystem. The technology can enhance the intelligence level of devices, achieve seamless user experiences, and support developers in innovating applications. Honor plans to promote the deep integration of AI with life, leading the development of intelligent lifestyles.

Zhipu launches the open-source large model GLM-4.5 series, available in two versions with 355 billion and 106 billion parameters, using a MoE architecture and following the MIT license. The model ranks second globally and first domestically in 12 benchmark tests, particularly excelling in tool calling and task completion. It supports complex tasks such as code generation and 3D interaction. API pricing is affordable (0.8 yuan per million tokens). Although the parameter count is only 1/2 to 1/3 of competitors, it achieves efficient performance through 15 trillion token pretraining.

Zhipu AI releases its next-generation flagship model, GLM-4.5, which achieves native integration of reasoning, coding, and agent capabilities. It uses a MoE architecture to provide high parameter efficiency. The model performs excellently across 12 authoritative evaluation benchmarks, achieving the third place globally and the first in domestic rankings. The model supports dual-mode operation, with training data reaching 2.3 trillion tokens, and offers competitive API pricing. GLM-4.5 is compatible with mainstream code frameworks, possessing full-stack development capabilities. It is now open-sourced on multiple platforms and provides an experience. Its technological breakthroughs and cost-effectiveness are expected to drive