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.

The Kimi Open Platform launched the Kimi K2 High-Speed Version (kimi-k2-turbo-preview), increasing the output speed from 10 Tokens per second to 40 Tokens per second while maintaining the original parameter scale, significantly improving response efficiency. This upgrade is mainly targeted at applications requiring higher real-time performance, aiming to optimize user experience.

kimi-k2-turbo-preview, the high-speed version of Kimi K2, is officially released. According to the information, this high-speed version model maintains the same parameter settings as the original kimi-k2, but achieves significant performance improvements. The output speed has jumped from 10 Tokens per second to 40 Tokens per second, providing users with a smoother and more efficient experience.

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

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.....

Pudong New Area launched a seed fund with a total scale of 2 billion yuan for artificial intelligence. The first phase is 500 million yuan, making it the first state-owned capital venture capital fund in China focusing on the pre-seed stage. It has signed 18 projects. The AI industry scale in Pudong exceeds 160 billion yuan, accounting for 40% of the city, and has gathered more than 200 specialized model enterprises. In the future, five major measures will be implemented: releasing over 30 application scenarios, building a heterogeneous robot training field, establishing a full-cycle fund matrix, optimizing industrial policies, and providing talent services. The goal is to add 1,000 AI companies within three years, and the core output value will exceed 200 billion yuan, aiming to build a comprehensive AI ecosystem.

VLM2Vec-V2, a multimodal embedding framework developed by Salesforce Research and other institutions, breaks through the limitations of traditional models. Based on the Qwen2-VL architecture, this framework innovatively unifies image, video, and visual document retrieval tasks. It adds five new evaluation tasks to expand the MMEB dataset. With key technologies such as dynamic resolution and M-RoPE, it achieved an average score of 58.0 across 78 datasets, leading in performance, especially in video tasks. Although its document retrieval is slightly behind ColPali, it

Zhiyuan Robotics announced the open-source plan for the Zhiyuan Lingqu OS, an embodied intelligence operating system, at the 2025 World Artificial Intelligence Conference. The humanoid robot Lingxi X2 demonstrated intelligent interaction capabilities through natural dialogue. The system adopts a layered open-source model, including underlying middleware, agent framework, and toolchain, with plans to launch in the fourth quarter. Zhiyuan Robotics has already achieved application in fields such as industrial logistics, and its Qiyuan general embodied large model won the SAIL Star Award. The company will continue to improve the operating system and promote the development of embodied intelligence industry.

SenseTime to launch Rixin v6.5, embodied AI platform, showcasing multi-modal & agent capabilities. Aims to integrate AI with physical world, boost productivity & interaction. Partners with ecosystem for AI industrialization & introduces 'Compute Mall' platform.....

Kuaishou opensources the KAT-V1 autonomous thinking large model, which includes two versions: 40B and 200B. The 40B version performance is close to DeepSeek-R1, and the 200B version outperforms several flagship models. The model innovatively adopts a mixed training paradigm of short and long-term thinking and the Step-SRPO reinforcement learning algorithm, which can automatically adjust the thinking mode based on the complexity of the question, solving the problem of overthinking. Based on Qwen2.5-32B, it achieves excellent performance in fields such as science and code through an heterogeneous distillation framework and pre-training with 10 million examples.

On July 23, Kunlun Group introduced two innovative AI models to global users - the Mureka V7 Music Model and the Mureka TTS V1 Audio Model. The launch of these two models marks an important step forward in the field of AI music generation. Since its release, the Mureka V7 Music Model has attracted widespread attention, gaining extensive praise for its efficient music creation capabilities and outstanding music quality. Users simply need to input lyrics, select a style or theme, and click generate to obtain a complete and pleasant song in a short period of time.

The Kimi team released the Kimi K2 technical report. This trillion-parameter large model, with 32 billion activated parameters, won the global open-source model competition, with performance comparable to top closed-source models such as GPT4.5. Its innovations include: using the MuonClip optimizer to process 15.5 trillion token data without loss; developing a multi-domain tool data synthesis pipeline; pioneering the restatement method to improve training efficiency, achieving the effect of ten rounds of original data in one round of rewritten data; optimizing model performance through supervised fine-tuning and reinforcement learning. The model is based on N

DeepSeek's Q2 downloads plunged 72.2% to 22.58M, losing market appeal. Doubao emerged as new leader while Kimi declined. Experts cite lost cost edge, cheap APIs, user churn, weak use cases, and R2 delays. Meanwhile, 'AI+Office' grew 190.2% and 'AI+Education' hit 10M users, signaling intense competition demanding constant innovation.....

In 2025, open-source large models will show three major trends: 1) MoE architecture becomes mainstream, with DeepSeek-V3 (67.1 billion parameters) and Qwen3-235B (235 billion parameters) each having unique designs in their expert systems; 2) small models break through performance bottlenecks, with SmolLM3-3B adopting position-encoding-free technology, and Qwen3-4B achieving lightweight efficiency; 3) models show significant differentiation, with Llama3.2 focusing on general tasks, while Kimi-K2 (1 trillion parameters) excels in complex reasoning.

Tesla has released its new Dojo 2 chip, with performance 10 times that of the first generation, and computing power close to NVIDIA B200. The chip is manufactured by TSMC and uses advanced packaging technology, solving power consumption issues. Dojo 2 will assist Tesla's FSD autonomous driving system training, processing 16 billion video frames per day, achieving technological self-sufficiency. Musk revealed that next year a more powerful Dojo 3 will be launched, and he joked that Dojo 2 can run 'Crysis' at one billion frames per second. This breakthrough will reduce Tesla's reliance on NVIDIA and may be offered externally.

Moonshot AI's open-source model Kimi K2 excels in programming, matching Qwen3-235B-A22B and nearing o3-mini-high & Claude-3.7-Sonnet. With 1T MoE, 128k context, 65.8% accuracy on SWE-bench, and $0.14/M tokens, it's ideal for coding. Supports web generation, workflows, vLLM/Hugging Face deployment, MIT-licensed.....

Suno AI V4.5+ introduces vocal/instrumental swap, playlist-inspired composition, 8-minute tracks, and enhanced audio for paid users.....