Z.ai 揭晓 Ox Alpha 实为 GLM-5.3-Flash
Finally, Z .ai revealed that Ox Alpha was actually GLM-5.3-Flash.
做推理优化和国产算力选型的同学必看，GLM-5.3-Flash 的稀疏+线性注意力组合和 IndexPool 设计值得深挖，赶紧去读技术细节。
Finally, Z .ai revealed that Ox Alpha was actually GLM-5.3-Flash. So that means over the last few days all those 100 tn tokens/day of stealth traffic capacity was running on Chinese AI chips, with tens of thousands of domestic accelerators behind the service. not an NVIDIA GPU cluster.
最终，Z.ai 透露 Ox Alpha 实际上是 GLM-5.3-Flash。这意味着在过去几天里，所有那些每天 100 万亿 token 的隐形流量容量都是在国产 AI 芯片上运行的，背后有数万个国产加速器支撑服务，而非 NVIDIA GPU 集群。
5.3-Flash beats GLM-5.2 at one-tenth the price with only 18B active parameters.
5.3-Flash 以十分之一的价格击败 GLM-5.2，且仅使用 18B 激活参数。
GLM-5.3-Flash. has 320B params in total, but only 18B are active during inference.
GLM-5.3-Flash 总共有 320B 参数，但推理时只有 18B 处于激活状态。
It also uses 45 layers instead of GLM-4.5's 92, cutting the amount of work required for each token.
它还使用 45 层而非 GLM-4.5 的 92 层，减少了每个 token 所需的工作量。
The benchmark jumps are large too: against GLM-5.2, DeepSWE rises from 46.2 to 63.4 and AutomationBench from 26.2 to 48.8.
基准测试的提升也很大：与 GLM-5.2 相比，DeepSWE 从 46.2 升至 63.4，AutomationBench 从 26.2 升至 48.8。
There is an architectural change as well, that cuts attention compute 3x and per-layer KV cache 4.4x versus GLM-5.3.
架构上也有变化，与 GLM-5.3 相比，注意力计算量减少 3 倍，每层 KV 缓存减少 4.4 倍。
The novelty is mainly in the combination: GLM-5.3-Flash uses linear attention for cheap state tracking, then sparse attention with a lightweight indexer to retrieve only the distant context worth revisiting, instead of repeatedly attending across the full 1M-token window. They also introduces IndexPool, which compresses four indexer key vectors into one, and says the combined design
新颖之处主要在于组合：GLM-5.3-Flash 使用线性注意力进行低成本状态跟踪，然后使用带有轻量级索引器的稀疏注意力，仅检索值得回顾的远距离上下文，而不是反复关注整个 1M token 窗口。他们还引入了 IndexPool，将四个索引器键向量压缩为一个，并表示这种组合设计
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力