作者对10种模型与调用方案的组合进行了系统性测试，以评估它们在构建交互式Web应用中的表现差异1。本次测试采用统一的Three.js任务，要求生成单页面应用，涵盖无人机、警示灯、发光跑道、体积雾效果和镜头路径等功能1。
测试中使用了包括GLM 5.3 Flash Max在内的多个模型1。为了便于比较，作者记录了各组合的关键性能指标，其中输入tokens包含了缓存输入部分，输出tokens则代表生成总数，包括整个推理过程1。值得注意的是，DSH适配器未单独报告推理计数，其持续时间合计为两个回合的活跃时间，不包含回合间的停顿1。
An author conducted a comprehensive evaluation of ten different model and harness combinations on an identical Three.js task to assess their performance variations in generating single-page interactive web applications 1. The testing employed a goal-oriented approach and involved multiple models including GLM 5.3 Flash Max, with measurements tracking input tokens, output tokens, inference processes, and execution time 1.
The task requirements called for building a single-page Three.js application featuring a drone, warning lights, illuminated runway, volumetric fog effects, and camera path animations 1. Input token counts incorporated cached inputs, while output token totals represented the complete generation including the reasoning process 1. The DSH adapter did not separately report inference counts, and its recorded duration reflected the combined active time across two rounds of interaction, excluding pauses between rounds 1.
评论
还没有评论，欢迎留下第一条。