智谱GLM-5.2开源百万上下文模型
推荐指数 69.0 NO. 013 · 2026.06.14
发布2026/06/13Score128Comments48
为什么值得看
智谱通过Z.ai发布GLM-5.2，支持100万token上下文窗口，下周将以MIT协议完全开源。在中美AI管制博弈背景下，中国开源模型正获得地缘政治层面的叙事红利，对需要长文本处理的RAG、法律文档分析等场景是重要选项。
媒体预览
编辑判断
百万上下文+MIT开源的组合直接对标Google Gemini 1.5 Pro和Moonshot Kimi，但后者均未完全开源权重。GLM-5.2的真正杀招不在长度本身，而是把之前只有闭源API才能用的长文本能力变成了可本地部署的选项。
对做企业知识库、法律合规审查的团队来说，这意味着可以彻底摆脱按token计费的焦虑。但需要注意：百万上下文的推理成本依然很高，实际落地时大概率需要结合滑动窗口或分层检索做工程优化，不能 naive 地直接塞满。
社区反馈
意见分歧 49 条评论
核心争论：发布时间是否刻意配合美国AI管制窗口以获取地缘政治叙事红利
Is it a coincidence that both MiniMax and Z.ai are releasing frontier open weights models right as the USG is trying to impose a cap on model capability offered to the public?
I would say yes. You think they were sitting on a release waiting for the right marketing moment?
I think it's a possibility, because labs trying to one-up each other is a fairly common phenomenon at this point. Previous Opus releases were immediately followed by GPT releases, for example. At some point the timing stops being a mere coincidence.