智谱开源 GLM-5.3-Flash（Ox Alpha）

2026 年 8 月 26 日晚，智谱公开了一个秘密：此前以匿名身份 Ox Alpha 在 OpenRouter、OpenCode 上线盲测的模型，真实身份是 GLM-5.3-Flash——320B 总参、18B 激活的 MoE，GLM-5 系列首个原生多模态模型，MIT 协议开源。

这场发布的故事其实从 8 月中旬就开始了。一个没有名字的模型出现在 OpenRouter 上，首日登顶、单日 token 量达平台历史峰值 4 倍，终结了 DeepSeek 连续 56 天霸榜。社区叫它「牛来」——名字里带着期待与调侃。围绕它的真实身份与算力来源，猜测持续发酵：是哪个实验室？用了谁的算力？为什么匿名？这些问题在揭晓前没有官方答案，社区用盲测数据自己投票。

揭晓后的信息量更大。官方称 GLM-5.3-Flash 的训练使用数十万亿词元，由约 10 万张国产芯片组成的集群支撑——这是国产芯片集群第一次公开支撑前沿模型训练。需要带着厂商自述的边界读，但方向是清楚的：算力自主从「战略叙事」变成「可核验的训练细节」。定价同样激进：约为 GLM-5.3 的十分之一、Claude Opus 4.8 的四十分之一，MIT 协议意味着可自由商用。

同一天，GLM-5.3 权重也正式开放下载（744B-A40B，采用新的 GLM-5.3 License）。两周前「安全评估完成后开源」的承诺如期兑现。两场发布叠在同一天：一个用匿名盲测制造悬念，一个用承诺兑现建立信任——智谱把「发布」本身变成了一种产品。

「匿名盲测 + 身份揭晓」的发布形态值得单独读。它把社区参与写进了发布流程：模型先以匿名身份接受真实用户检验，再揭晓身份、公开细节。对开发者，盲测数据比厂商自报基准更接近真实使用体验；对行业，这种形态让「发布」从单向宣告变成双向对话——其他厂商会跟进，还是坚持「正式发布」的传统？

GLM-5.3-Flash 的意义不在单个数字。它把「匿名盲测 + 身份揭晓」变成一次完整的发布叙事，证明国产芯片集群可以支撑前沿模型训练；同时以 MIT 协议与极低定价，把开源旗舰的竞争从参数推向生态与成本。对开发者，这是一个可自由商用的原生多模态开源模型，成本约为闭源旗舰的四十分之一；对行业，开源与算力自主的叙事合流，匿名盲测成为模型发布的新形态。牛来了，牛走了，留下的是被改写的发布规则。

On the evening of August 26, 2026, Zhipu revealed a secret: the model that had been running anonymously as Ox Alpha on OpenRouter and OpenCode was GLM-5.3-Flash—a 320B-total, 18B-active MoE, the GLM-5 line's first native multimodal model, open under MIT.

The story of this launch actually began in mid-August. A model with no name appeared on OpenRouter, topped it on day one, hit 4x the platform's historical daily token peak, and ended DeepSeek's 56-day streak. Communities nicknamed it "牛来"—the name carrying both anticipation and teasing. Speculation about its identity and compute raged: which lab? whose chips? why anonymous? There were no official answers before the unmasking, so the community voted with blind-test data.

The reveal carried more. Zhipu said GLM-5.3-Flash was trained on tens of trillions of tokens on a cluster of roughly 100,000 domestic chips—the first time a domestic-chip cluster publicly supported frontier model training. Read with the vendor-self-reported caveat, but the direction is clear: compute sovereignty moved from "strategic narrative" to "verifiable training detail." Pricing is equally aggressive: about one-tenth of GLM-5.3 and one-fortieth of Claude Opus 4.8, under MIT for free commercial use.

The same day, GLM-5.3 weights (744B-A40B) also opened under the new GLM-5.3 License. The "open after safety evaluation" promise from two weeks earlier was kept on schedule. Two launches stacked on one day: one built suspense through anonymous blind testing, the other built trust through a kept promise—Zhipu turned "launching" itself into a product.

The "anonymous blind test plus unmasking" format deserves its own reading. It wrote community participation into the release process: a model first faces real users anonymously, then its identity and details are revealed. For developers, blind-test data is closer to real usage than vendor-reported benchmarks; for the industry, this format turns "launch" from a one-way announcement into a two-way conversation—will others follow, or hold to the "formal release" tradition?

GLM-5.3-Flash's meaning is not in any single number. It turned "anonymous blind test plus unmasking" into a complete launch narrative, showing a domestic-chip cluster can support frontier training; with the MIT license and rock-bottom pricing it also pushed open-flagship competition from parameters toward ecosystem and cost. For developers, this is a freely commercial native multimodal open model at about one-fortieth of closed-flagship cost; for the industry, open-source and compute-sovereignty narratives merged, and anonymous blind testing became a new launch format. The ox came, the ox went, and what remains is a rewritten rulebook for launches.