[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"aihot-art-90180":3},{"itemId":4,"vertical":5,"category":6,"source":7,"score":8,"title":9,"summary":10,"analysis":11,"url":12,"coverUrl":13,"direction":13,"marketSignal":13,"publishedAt":14},"90180","ai","模型发布\u002F更新","IT之家",58,"腾讯发布具身VLM基座模型Hy-Embodied-VLM-1.0，3B规模性能接近上代32B模型","腾讯发布具身VLM基座模型Hy-Embodied-VLM-1.0,3B参数性能接近上代32B模型。","腾讯Robotics X实验室联合福田实验室与腾讯混元发布第二代具身VLM基座模型Hy-Embodied-VLM-1.0,只用3B(A3B)规模就把整体性能做到接近上一代32B(A32B)旗舰模型,是具身智能基座模型“小型化”的一个信号。\n\n· 在覆盖37个评测任务的具身能力体系中,模型在物理状态理解、动作变化推理、时序与自适应推理三个维度分别拿到68.6、64.1、57.4分,综合均分65.6\n· 同等3B规模下,明显优于Qwen3.6-A3B、Cosmos 3-8B、Embodied-R1等通用与具身模型\n· 能力构建分三层:物理空间状态理解(识别物体、深度、可操作区域)、动作与变化理解(下一步动作、轨迹定位、可执行性判断)、长时程时序与自适应推理(多步规划、视觉语言导航、失败诊断、动态重规划)\n· 模型权重已在GitHub与Hugging Face开源,可直接下载使用\n\n对机器人和具身智能创业团队而言,一个3B规模就能打平上代32B性能的基座模型,意味着具身VLM的部署门槛和推理成本会显著下降。","https:\u002F\u002Fwww.ithome.com\u002F0\u002F978\u002F080.htm",null,"2026-07-17 15:22:18"]