智谱发布 GLM-5.3

2026 年 8 月 14 日，智谱发布 GLM-5.3。发布材料里最值得注意的不是参数——基座与 GLM-5.2 相同——而是两件事：官方称编程能力较前代提升约 50%，以及「安全加固与评估完成后，两周内开源权重」的承诺。

先看第一件。在 Kimi K3 的 2.8T 与 Qwen3.8-Max 的 2.4T 把开源旗舰推向规模竞赛时，智谱选择了一条不同的路：基座不变，用极致的后训练 Scaling（强化学习）提升智能上界。官方口径里，GLM-5.3 是「当时公认的编程能力最强开源模型」，在 Terminal Bench 3.0、Agents' Last Exam（CLI）等基准取得开源第一。这些数字需要带着厂商自述的边界读，但方向是清楚的：规模不是唯一的路，后训练质量可以改写智能上界。

第二件更值得读。GLM-5.3 采用新的 GLM-5.3 License，取代前代的 MIT：年营业额超 100 亿美元的机构，若将其作为外部模型服务对外提供，须先通过智谱安全审查。这是开源模型第一次把「安全审查」写进许可条款——开源的定义与边界被重新讨论。支持者认为这是负责任的开源，批评者担心它开了「有条件开源」的先例。无论立场如何，一个事实是确定的：开源不再只是「权重放出来」，而是「在什么条件下放出来」。

「两周后开源」的承诺同样值得注意。它把开源从发布后的可选项变成发布流程的一部分——先发布、再评估、后开源，节奏被写进产品计划。这与智谱 8 月 3 日 GLM-5 预告时展现的「先预告、再发布、快速迭代」节奏一脉相承。

GLM-5.3 的意义不在单个数字。它证明在规模竞赛之外，后训练 Scaling 也能改写开源旗舰的智能上界；同时用「两周后开源」与新的许可条款，把安全审查写进开源流程。对开发者，这是一个编程能力领先的开源基座，两周后可本地部署与微调，但商用前需要读一遍许可；对行业，开源旗舰的竞争从参数规模延伸到后训练质量，安全审查成为开源流程的一部分。规模竞赛还在继续，但至少这一周，智谱展示了另一条路。

On August 14, 2026, Zhipu released GLM-5.3. The most notable things in the release materials are not the parameters—the base is the same as GLM-5.2—but two others: coding reported up about 50% from the previous generation, and a promise to open weights "within two weeks after safety hardening and evaluation."

On the first: while Kimi K3's 2.8T and Qwen3.8-Max's 2.4T pushed open flagships into a scale race, Zhipu chose a different road—same base, intelligence ceiling raised through aggressive post-training scaling (RL). By Zhipu's account, GLM-5.3 was "widely regarded as the strongest open coding model at the time," taking first place among open models on Terminal Bench 3.0 and Agents' Last Exam (CLI). Read with the vendor-self-reported caveat, but the direction is clear: scale is not the only road; post-training quality can rewrite the intelligence ceiling.

The second deserves more attention. GLM-5.3 ships under a new GLM-5.3 License replacing MIT: institutions with over $10 billion in annual revenue must pass Zhipu's safety review before offering it as an external model service. This is the first time an open model wrote "safety review" into its license terms—reopening the question of what open means. Supporters call it responsible open source; critics worry it sets a precedent for "conditional open." Either way, one fact stands: open is no longer just "weights released" but "released under what conditions."

The "open in two weeks" promise matters too. It turns open-sourcing from an optional post-launch step into part of the release process—release first, evaluate, then open, with the cadence written into the product plan. This continues the rhythm Zhipu showed with the GLM-5 teaser on August 3: tease first, release, iterate fast.

GLM-5.3's meaning is not in any single number. It showed that post-training scaling can rewrite the open flagship's intelligence ceiling outside the scale race; the "open in two weeks" promise and the new license also wrote safety review into the open-source process. For developers, this is a leading open coding base, deployable and fine-tunable in two weeks—but read the license before commercial use; for the industry, open-flagship competition extended from parameter scale to post-training quality, with safety review entering the open-source process. The scale race continues, but this week at least, Zhipu showed another road.