Primary sourceshao__meng
View sourceEvent ·Multi-source
SemiAnalysis 播客拆解 GLM 5.3:后训练提升与国产芯片全栈适配
SemiAnalysis 播客 Ep. 038 围绕 Z AI 两个月前发布的开源模型 GLM 5.3 做技术拆解。5.3 与 5.2 架构完全相同,agentic 基准提升超过 2%,全部来自基于 GRPO 的分阶段后训练,并开源了 slime 后训练框架。InferenceX 基准显示 AMD MI355 凭借 ATOM 推理引擎的迭代,几周后在每百万 token 服务成本上反超 NVIDIA B200。GLM 5 论文写明模型从第一天起全栈适配华为昇腾、摩尔线程、寒武纪等七大国产芯片平台,5.3 Flash 曾匿名以 "aux alpha" 之名在 OpenRouter 上完全由中国硬件提供服务。网络安全能力从 5.2 到 5.3 被形容为昼夜之差,Anthropic 也承认其接近 Fable 级水平。
Current read
SemiAnalysis 这期播客把 GLM 5.3 拆得很细:架构没动全靠后训练,还从第一天就适配七大国产芯片,安全攻防圈都在拿它替代闭源模型。
11 sources79° AI heatLast updated 2026/10/18 13:43:31
Evidence
11 sourcesRelated source 1NVIDIA AI
View sourceRelated source 2TestingCatalog
View sourceRelated source 3berryxia
View sourceRelated source 4kimmonismus
View sourceRelated source 5The Rundown AI
View sourceRelated source 6LMSYS Org (SGLang)
View sourceRelated source 7Teortaxes
View sourceRelated source 8Jim Fan
View sourceRelated source 9Epoch AI
View sourceRelated source 10vLLM
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.