???ڇ?;

Chinese AI models sweep top spots in major rankings as open source, cost advantages drive overseas adoption despite US tech restrictions: expert_我的网站

A |     IT之家 7 月 21 日消息，科技媒体 The Decoder 昨日（7 月 20 日）发布博文，报道称谷歌 DeepMind 发布 GenCeption 模型，将预训练的视频生成器重新用于深度估计和分割等经典计算机视觉任务。

B |          IT之家援引博文介绍，大语言模型在学习预测下一个 Token 的时候，在训练过程中往往需要吸收语法、世界知识和上下文关系等内容。    

    Conceptual diagram of AI Photo: VCG
Chinese artificial intelligence (AI) technology companies are gaining a growing foothold among international users, with their models topping rankings on some leading industry platforms, ahead of their US counterparts.
Chinese AI companies, with their open-source strategies and cost effectiveness, are gaining growing adoption among overseas developers despite the continued US restrictions on China's access to advanced chips and AI technologies, a Chinese veteran industry analyst said, noting that global AI competition should be about innovation capabilities and the ability to build attractive ecosystems, rather than about who can erect higher technological barriers.
The latest weekly large-language model (LLM) rankings from AI model aggregation platform OpenRouter showed that models developed by Chinese companies swept the top four spots and took six of the top 10, the Global Times noted on the platform on Tuesday. DeepSeek V4 Flash 0731 topped the ranking with 9.39 trillion tokens in usage, followed closely by Tencent's Hy3 with 8.94 trillion.
From August 3 to Sunday, Chinese AI models recorded 34.25 trillion tokens in usage, up 21.76 percent from the previous week, while US AI models, although up 109.36 percent, only recorded 9.17 trillion tokens, about three times less than that of Chinese models.
Another DeepSeek model, V4 Flash 0423, came in third, followed by Xiaomi's MiMo-V2.5 in fourth place. Zhipu AI's GLM 5.2 and DeepSeek V4 Pro also made the top 10.
Notably, usage of the top-ranked DeepSeek V4 Flash 0731 surged 302 percent over the week. The rankings are based on actual model usage on OpenRouter, highlighting the growing adoption of Chinese AI models among global developers and their strong presence on the platform, Xiang Ligang, director-general of the Zhongguancun Modern Information Consumer Application Industry Technology Alliance, told the Global Times on Tuesday.
Meanwhile, the latest trending rankings on Hugging Face, the world's largest open-source AI community, also highlighted the growing influence of Chinese AI models among global developers. As of Tuesday, four of the top five models on the platform's Trending Models list were linked to Chinese AI companies, with MiniMax's newly released full modality model MiniMax H3 topping the ranking.
One key reason Chinese AI models are gaining popularity is their combination of strong performance, lower costs and greater openness, Chinese experts said.
"Companies such as DeepSeek, MiniMax and Zhipu AI have released open-weight models that developers can download, customize and use to build their own applications. Many Chinese models also come with relatively low API and inference costs, making them more affordable for start-ups and smaller businesses," Xiang said.
According to UBS estimates cited by media reports, leading Chinese AI models cost about one-tenth as much to train as comparable overseas models, while their API prices are typically only 10 percent to 20 percent of those of foreign alternatives.
"China's strong industrial foundation in power supply, computing infrastructure and data center development also provides solid support for the large-scale deployment of AI," said Xiang.
In contrast to many large US AI companies that primarily offer their flagship models through closed platforms and application programming interfaces, China's are gaining traction among global developers with its open model approach, Ma Jihua, a veteran industry analyst, told the Global Times on Tuesday.
The open model approach allows a much broader range of developers, including smaller players, to participate in AI development rather than leaving advanced AI capabilities concentrated in the hands of a few large companies, said Ma. As more developers build on these models, they can create new tools and applications, further expanding the ecosystem around Chinese AI models, Ma added.
China's recently released Action Plan for Global AI Governance explicitly encourages the joint development of international open-source AI communities, greater international exchanges and cooperation among open-source communities, and the sharing of general purpose large models, foundational algorithms and software tools.
Recently, the US government has ramped up its unilateral moves targeting China's AI-related sector, including a plan to ban cutting-edge Chinese AI models, US media outlets have reported.
The growing influence of Chinese AI models shows that technological blockades cannot hold back innovation and industrial development, Ma said, noting that such restrictions could instead backfire by raising costs for US companies and limiting their access to competitive technologies and global AI cooperation.
Chinese AI models are already seeing growing adoption among US businesses. Chinese AI models now consistently account for more than 30 percent of weekly token usage by US companies, with the share peaking at 46 percent, the People's Daily reported, citing data from a third party industry platform.
The US should view the development of China's AI industry objectively, and avoid politicizing or overstretching the concept of national security in normal technological and industrial competition, Ma said. 
"After all, global AI competition should ultimately be about innovation and the ability to build attractive ecosystems, rather than about who can erect higher technological barriers," Ma noted.
。         但是在计算机视觉领域，视觉模型缺少等效的训练方法，主要由专业模型主导，包括用于分割的“Segment Anything”和用于深度估计的“Depth Anything”，每个模型都使用其特定的架构。         谷歌 DeepMind 团队为此提出 GenCeption 模型方案，尝试将一个“生成视频”的 AI 模型逆向改造成一个能“理解世界”的视觉分析引擎。         GenCeption 打破了传统计算机视觉“一个任务一个专用模型”的格局，仅凭单一模型就能同时做好深度估计、图像分割、3D 姿态估计、表面法线预测和相机姿态估计等核心视觉任务。         GenCeption 基于阿里巴巴开源视频模型通义万相 Wan2.1 系列训练，与传统扩散模型需多步去噪不同，GenCeption 在一次前向传播中完成预测，从而提升视觉任务处理速度。模型通过文本提示指定任务，可输出深度图、表面法线图、分割掩码，并可处理相机运动表示。

C |          训练数据以合成为主。论文称，数据集仅包含 7500 段视频，由 800 个数字人体模型与 200 段动作捕捉序列组合生成，再通过 Blender 在不同背景和镜头角度下渲染。         泛化方面，GenCeption 几乎只在单人合成视频上训练，但可处理真实多人视频，也可迁移到动物和类人机器人类别。论文称，部分输出细节甚至超过训练时 Blender 渲染结果，可保留猫胡须和单根发丝边缘。         性能方面，论文给出两组处理时间数据：小模型处理 81 帧视频约需 6 秒；大模型参数量为 140 亿，处理同样长度视频约需 10 秒。         参考。

Current article:http://j95.dinggaxilianmaiyuejienianguai.shop/b3daelv/20260826/1302404.html

Published on:22:58:48