微软推出两款全新人工智能编码模型MAI Code 2和MAI Code 2 Flash,旨在以远低于竞争对手的成本提供企业级编码任务的高性能 1。其中,旗舰模型MAI Code 2在DeepSWE 1.1等四项基准测试的平均得分中达到了GPT 6.1 Sol性能的98%,而成本仅为其40%;相较于Claude Opus 5.5,其性能达到96%,成本则仅为后者的5% 1。此外,在盲测前端应用评估环节中,有67%的情况显示评审员认为MAI Code 2的表现优于GPT-6 Sol 1。
针对高频日常任务优化的MAI Code 2 Flash则在三个高频编码基准测试中展现了极具竞争力的性价比,其性能与GPT-6 Luna相当但成本为后者65%,或与Claude Haiku 5.5相当但成本仅为后者20% 1。基于Flash架构衍生的MAI Work Flash已集成至Excel、PowerPoint等Microsoft 365核心产品中,并在多项内部基准测试中表现优异:在Word应用中,它以比GPT 6.1 Sol少约42%的输入令牌实现了接近的效果;在Teams应用中,它比GPT 6 Luna少用28%的令牌且效果更佳;而在FabricIQ中,其DAX查询生成能力超越了GPT-5.6-sol等模型 1。目前,MAI Code 2和Code 2 Flash已在GitHub Copilot上线,Code 2 Flash同时在Microsoft Foundry开启有限预览,后续将上线Open Router、Vercel等平台 1。
Microsoft has released two new artificial intelligence coding models, MAI Code 2 and MAI Code 2 Flash, designed to deliver high-performance enterprise coding tasks at significantly lower costs than competitors 1. The flagship model, MAI Code 2, achieves 98% of GPT-6.1 Sol’s performance across four benchmarks (including DeepSWE 1.1) while costing only 40% as much, and reaches 96% of Claude Opus 5.5’s capability at just 5% of its cost 1. In blind evaluations of frontend applications, reviewers preferred MAI Code 2 over GPT-6 Sol in 67% of cases 1.
MAI Code 2 Flash targets high-frequency daily tasks, matching the performance of GPT-6 Luna at 65% of the cost and Claude Haiku 5.5 at 20% of the cost across three coding benchmarks 1. A derivative model optimized from Flash, known as MAI Work Flash, has been integrated into core Microsoft 365 products including Excel and PowerPoint 1. This integration yields specific efficiency gains: in Word, it uses approximately 42% fewer input tokens to achieve results near those of GPT-6.1 Sol; in Teams, it requires 28% fewer tokens than GPT-6 Luna while delivering better outcomes; and in FabricIQ, it surpasses models like GPT-5.6-sol in DAX query generation 1. Both MAI Code 2 and Code 2 Flash are currently available on GitHub Copilot, with Code 2 Flash also in limited preview on Microsoft Foundry, and future availability is planned for platforms such as Open Router and Vercel 1.
评论
还没有评论,欢迎留下第一条。