[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fauDBX46gccUSmpNEm1fmn3fb9aMOK6uqpXyZrrvQJJM":3},{"code":4,"msg":5,"data":6},200,"操作成功",{"id":7,"title":8,"content":9,"digest":10,"source":10,"coverPath":11,"thumbsCoverPath":12,"isTop":13,"isShow":14,"baseClick":13,"clickCount":15,"createTime":16,"typeId":17,"isNewest":18,"newsInfoTypeRespVo":19,"voiceUrl":22,"voiceSize":23,"taskId":24,"releaseTime":25,"titleEn":26,"contentEn":27,"voiceUrlEn":28,"taskIdEn":29,"voiceSizeEn":30},1378,"Meta 推出小型推理模型MobileLLM-R1，企业应用向 “小型 AI” 转型","\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">近期，Meta 公司推出了一款名为 MobileLLM-R1的小型推理模型，引发了业界对 “小型 AI” 在企业应用中的关注。以往，人工智能模型的强大能力往往与其庞大的参数规模相关，许多模型的参数量达到数百亿甚至数万亿。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">然而，超大规模模型在企业使用时存在诸多问题，例如对底层系统缺乏控制、依赖第三方云服务以及成本不可预测等。针对这些痛点，小型语言模型（SLMs）的发展势头逐渐增强，旨在满足企业对成本、隐私和控制的需求。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002F78c9db809187455ba3a8299bf970593d\u002F6389381149364285646408021.png\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">MobileLLM-R1系列模型包含140M、360M 和950M 三种参数规模，专门针对数学、编码和科学推理进行优化。这些模型采用了 “深而薄” 的架构设计，通过优化的训练过程，使其在资源受限的设备上能够执行复杂任务。此外，MobileLLM-R1在 MATH 基准测试中的表现略优于阿里巴巴的 Qwen3-0.6B，尤其在 LiveCodeBench 编码测试中更是表现出色，适合在开发工具中进行本地代码协助。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">值得注意的是，MobileLLM-R1目前仅在 Meta 的 FAIR 非商业许可下发布，禁止任何商业用途，因此更适合作为研究蓝图或内部工具，而非可以直接用于商业化的产品。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在小型语言模型的竞争环境中，Google 的 Gemma3（270M 参数）以其超高效的性能而闻名，且其许可证更为宽松，适合企业进行定制。与此同时，阿里巴巴的 Qwen3-0.6B 也是一个很好的选择，提供了不受限制的商业使用。Nvidia 的 Nemotron-Nano 则在控制功能上具有独特优势，支持开发者根据需求调整推理过程。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">随着企业逐渐认识到小型模型带来的可控性和经济性，行业正经历向小型专用模型的转型。许多企业意识到，小型模型能够提供更高的可预测性和隐私保护。此外，利用一系列小型专用模型来解决复杂问题的思路，也与软件行业向微服务架构的转变相似。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">这一变化并不意味着大型模型将会被淘汰，相反，它们将继续发挥作用，通过优化训练数据，为新一代小型模型提供理想的训练集。这一趋势表明，未来的 AI 发展将更加可持续，各大公司正朝着更加务实的 AI 未来迈进。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"color: rgb(187, 187, 187);\">【新闻来源】AIbase基地 \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.aibase.com\u002Fzh\u002Fnews\u002F21398\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">https:\u002F\u002Fwww.aibase.com\u002Fzh\u002Fnews\u002F21398\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">（本网转发此文章，旨在为读者提供更多的信息资讯，所涉内容不构成投资、消费建议。文章事实如有疑问，请与有关方核实，文章观点非本网观点，仅供读者参考。）\u003C\u002Fspan>\u003C\u002Fp>","","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002Fd2ff6b5fd4a24903ac380dfa109940a7\u002FAI领域.jpg","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002Fthumbs\u002Fd2ff6b5fd4a24903ac380dfa109940a7\u002FAI领域.jpg",0,1,56,"2025-09-19 18:38",2,false,{"id":17,"name":20,"enName":21},"芯位视野","Xinwei Vision","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3Aa97f65b0-ef95-47fe-8081-30031000dccd%3A0.wav?Expires=1758282056&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=j%2FSeYs9YQRwE7zwb%2FB3BHp%2F5bkI%3D",4695162,"a97f65b0-ef95-47fe-8081-30031000dccd","2025-09-19 18:36","Meta launches small reasoning model MobileLLM-R1, enterprises' transformation towards \"small AI\"","\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">Recently, Meta launched a small reasoning model called MobileLLM-R1, which has drawn attention from the industry regarding the use of \"small AI\" in enterprise applications. In the past, the powerful capabilities of artificial intelligence models were often related to their large parameter scale, with many models having hundreds of billions or even trillions of parameters.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">However, large-scale models have many issues when used by enterprises, such as lack of control over the underlying system, reliance on third-party cloud services, and unpredictable costs. To address these pain points, the development of small language models (SLMs) is gradually gaining momentum, aiming to meet enterprises' needs for cost, privacy, and control.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002F78c9db809187455ba3a8299bf970593d\u002F6389381149364285646408021.png\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The MobileLLM-R1 series of models includes three parameter scales: 140M, 360M, and 950M, specifically optimized for mathematics, coding, and scientific reasoning. These models use a \"deep and thin\" architecture design, enabling them to perform complex tasks on resource-constrained devices through an optimized training process. Additionally, MobileLLM-R1 performs slightly better than Alibaba's Qwen3-0.6B on the MATH benchmark test, and especially excels in the LiveCodeBench coding test, making it suitable for local code assistance in development tools.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Notably, MobileLLM-R1 is currently released under Meta's FAIR non-commercial license, prohibiting any commercial use, so it is more suitable as a research blueprint or internal tool rather than a product directly usable for commercial purposes.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In the competitive environment of small language models, Google's Gemma3 (270M parameters) is known for its ultra-efficient performance, and its license is more flexible, suitable for enterprise customization. At the same time, Alibaba's Qwen3-0.6B is also a good choice, offering unrestricted commercial use. Nvidia's Nemotron-Nano has unique advantages in control functions, supporting developers to adjust the reasoning process according to their needs.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">As enterprises gradually recognize the controllability and cost-effectiveness brought by small models, the industry is experiencing a transition towards small specialized models. Many enterprises realize that small models can provide higher predictability and privacy protection. Furthermore, the idea of using a series of small specialized models to solve complex problems is similar to the software industry's shift toward microservices architecture.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">This change does not mean that large models will be eliminated; instead, they will continue to play a role by optimizing training data to provide ideal training sets for the next generation of small models. This trend indicates that future AI development will be more sustainable, and major companies are moving toward a more practical AI future.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"color: rgb(187, 187, 187);\">【News source】AIbase Base \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.aibase.com\u002Fzh\u002Fnews\u002F21398\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">https:\u002F\u002Fwww.aibase.com\u002Fzh\u002Fnews\u002F21398\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">（This article is forwarded by this site to provide readers with more information, and the content does not constitute investment or consumption advice. If there are questions about the facts of the article, please verify with the relevant parties. The views of the article are not the views of this site, and are for reference only.）\u003C\u002Fspan>\u003C\u002Fp>","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3A2798249f-6e74-4f29-9904-520ba8322719%3A0.wav?Expires=1774838467&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=xPsh4jXgY0G2Mbbd6bfmepDLnoM%3D","2798249f-6e74-4f29-9904-520ba8322719",5743376]