微软完成MAI-Base-2预训练，总参数达28000亿
微软10月15日表示，已完成MAI-Base-2的预训练。这是一个混合专家模型，总参数28000亿，激活参数800亿，公司计划把权重授权给合作伙伴。
王雅琴人工智能、模型与技术
微软造了什么
8月和9月，微软用16000块英伟达GB200 GPU训练MAI-Base-2，历时38天。这次训练处理了50万亿token的文本和视觉数据。模型总参数28000亿，但每次输入只激活800亿。混合专家结构让它能保留一个大模型，又不必在每次运算中跑完全部参数。
与今年早些时候推出的MAI-Base-1相比，微软把独有的训练数据扩充了约36%。图像占新训练数据的18%，微软把这部分称为原生视觉训练，用于图表解读、文档分析和空间推理。微软称，模型的学习来自原生训练数据，而不是从其他AI模型蒸馏知识。
效率说法及其边界
微软估计，在选定测试上达到同等表现时，其训练方法的计算效率是DeepSeek-V4-Pro-Base的4.2到12.3倍。公司称MAI-Base-2在内部评测中胜过其中包含的开源权重模型。TechRepublic指出，这些数字估算的是达到可比表现所需的算力，并不说明MAI-Base-2全面占优。
这些比较依据的是微软自己挑选的留出评测，公司没有公布针对每一个领先闭源模型的同类对比。Business Insider报道称，微软AI负责人Mustafa Suleyman告诉员工，公司已在私下把MAI-Base-2与OpenAI的模型做过比较。微软没有公开这些结果，外界无法独立评估。
授权，而非公开发布
微软计划通过Microsoft Foundry把MAI-Base-2的权重授权给可信合作伙伴，而不是公开发布。微软尚未公布定价或正式可用时间。TechRepublic指出，获取条件和定制成本将影响企业是否选择微软的方案，而不是真正开源的替代品。
模型仍在进行后训练。报道还提到，企业不应指望马上通过Copilot或其他微软服务用上MAI-Base-2。微软表示，MAI-Base-3的准备工作已经启动。
TechRepublic报道称，该模型的文本和视觉能力最终可能支撑涉及文档、电子表格和图表的商业应用，但微软尚未展示这些能力如何变成商业产品。
微软为什么要自研模型
据eWeek报道，微软已向OpenAI投资约130亿美元，并继续在要求更高的任务上使用OpenAI的模型。公司也开始在一些产品中使用自研的MAI模型以降低成本。据彭博社报道，Suleyman称MAI模型已部署在微软一半以上的产品中，目标是减少并最终取消微软向Anthropic支付的费用。
据Yahoo Finance报道，微软在4月修订了与OpenAI的协议，结束了对OpenAI知识产权的独家使用权，并取消了微软向OpenAI的付款，而OpenAI继续向微软付款，但设有上限。据eWeek报道，微软仍为OpenAI的模型提供算力基础设施。
资料来源
32- 01Microsoft Unveils 2.8 Trillion-Parameter AI Model, Claims Major Efficiency GainsEN
- 02Microsoft's In-House AI Replaces OpenAI Models in PowerPoint, Bing and OneDriveEN
- 03Microsoft Builds In-House AI to Cut Costs Across Office AppsEN
- 04Microsoft debuts in-house AI models as it looks to ease reliance on OpenAIEN
- 05Microsoft CEO Satya Nadella shares MAI AI strategy, says new models will cut costs and power Copilot, Outlook, ExcelEN
- 06The Two-Ledger Model: Transaction-Grain Cost Accounting for Multi-Agent AI Systems Across Clouds, Models, and Systems of RecordEN
- 07Open models can do 90 percent of agentic work at a twentieth of the costEN
- 08Protesters Break Into Microsoft Data Center in AmsterdamEN
- 09Microsoft ist kein Schicksal: Warum Europa jetzt handeln mussEN
- 10Sodium-ion batteries approach cost parity with LFP, with 70 GWh covered by supply agreementsEN
- 11Build World Models for Interactive Simulation with Nvidia CosmosEN
- 12New Nvidia Cosmos 3 Open Models for Interactive World SimulationEN
- 13For host Same model, 169x smaller prompt: cost per correct step fell 4.5xEN
- 14 Boost your Nintendo Switch 2's storage for Zelda: Ocarina of Time with these microSD Express deals, starting at $48 EN
- 15Fitting AI Models To Edge ApplicationsEN
- 16Microsoft Paint RCE Fix Now Available, but Windows Update Won't Deliver ItEN
- 17When Bigger Becomes Brittle: Why Compression Hurts More as Language Models ScaleEN
- 18Microsoft admits certifications can be a drag, launches creds that recognize skillsEN
- 19Synthesizing symbolic models of human and machine reasoning at scaleEN
- 20SAVeR$^2$: Reasoning-based Safety Alignment for Large Reasoning Models via Verifiable RewardsEN
- 21GuidedWAM: Training-Free Energy Guidance for World Action Models in Autonomous DrivingEN
- 22Big Models, Small Gains: Frozen Forecasters as Classifiers on MONSTEREN
- 23Stochastic Dynamics of Generative AI: From Autoregressive Models to Interacting LLM Networks and Neural SDEsEN
- 24Shaping & Monitoring Models through their Internal RepresentationsEN
- 25Compositional Unalignment: Auditing Safety in Large Vision-Language ModelsEN
- 26Adversarial Multiplicity: Cherry-Picked Models and the Limits of DisclosureEN
- 27Fake OpenAI, SpaceX events used to steal Microsoft, Google accountsEN
- 28The Wizard of Open Weights: Chris Alexiuk Makes AI Easy With Open ModelsEN
- 29Anthropic's FlashProve proves which model wrote a reply, and its 72x costEN
- 30As Diesel Costs Surge for Farmers & Rural Businesses, Solar Offers a LifelineEN
- 31New Research Tests Whether AI Models Know Their Own ThoughtsEN
- 32Launch HN: Rivet (YC W23) – Scale and Orchestrate Agents on S3EN
本文所有数字与引语均来自下列来源，未添加来源之外的内容。
本内容由编辑部在人工智能辅助下制作完成。
王雅琴
评论
0- 还没有评论——来说两句吧。