NVIDIA marks its hosted GLM-5.3 API for deprecation: October 27 notice, support ends after October 28
NVIDIA's model page for z-ai/glm-5.3 now carries a deprecation banner: the API is deprecated on October 27, 2026 and unsupported after October 28. This covers NVIDIA's own hosted trial endpoint, not Z.ai's service and not the model itself. What the page says verbatim, what it does not mean, and what to check in your code.
NVIDIA has put a deprecation banner on its hosted GLM-5.3 model page. The notice is short and specific: the API is deprecated on October 27, 2026 and will no longer be supported after October 28, and users should transition to another model. Because GLM-5.3 is a model people reach through several different doors - Z.ai's own API, third-party hosts, open weights - the banner needs careful reading before it becomes a headline. This is what the page actually says, checked today, and the boundaries of what it means.
The short version
NVIDIA's build page for the model z-ai/glm-5.3 now opens with a notice: "This API will be deprecated on 10/27/2026. It will no longer be supported after 10/28/2026. Please transition to another model to avoid any service interruptions." The page gives no exact cutoff time or timezone, and it names no replacement model. The notice covers NVIDIA's own hosted endpoint - the trial service reachable through the base URL integrate.api.nvidia.com/v1 with the model name z-ai/glm-5.3. It does not say anything about Z.ai's own API, about GLM-5.3-Flash, about other providers hosting the same model, or about the model's open weights. If your code calls that exact endpoint with that exact model string, you have a migration to plan this week; everyone else should file this under "hosted trial endpoints are temporary by design."
What the page says, verbatim
Two NVIDIA pages tell the story. The model's page on NVIDIA's build site carries the deprecation notice quoted above, and as of today it still lists "Free Endpoint: Available" alongside "Partner Endpoint: Available." The reference documentation for the same model describes it as a trial service governed by the NVIDIA API Trial Terms of Service. In other words, the thing being retired is NVIDIA's hosted access path to the model - a prototype endpoint that was always labeled a trial. The vendor's entire migration guidance is the sentence asking users to transition to another model and to consult the API Reference for alternatives. No successor model is named, and no automatic migration is promised.
What this does not mean
Precision matters here, because "GLM-5.3 deprecated" is the kind of phrase that travels badly. The banner does not say GLM-5.3 the model is being withdrawn. It does not say Z.ai's own API is affected - that is a separate service run by the model's developer, and nothing on NVIDIA's page speaks for it. It does not mention GLM-5.3-Flash, the lighter sibling model, at all. It does not touch the model's weights or its license: the same documentation describes GLM-5.3 as released under an MIT-like license with an extra security-review clause for very large model-as-a-service operators, and a deprecation of one hosted endpoint changes none of that. And the page still lists a partner endpoint as available, so even within NVIDIA's ecosystem the wording is "this access path closes," not "this model vanishes."
The model in one paragraph
GLM-5.3 is Z.ai's 753-billion-parameter mixture-of-experts text model, built on the same base as GLM-5.2 with gains coming from post-training, aimed at complex coding and long-horizon agentic work. It uses DeepSeek-style sparse attention over a context of about one million tokens, ships with native FP8 weights, and supports reasoning and tool calling. Those are the vendor's own specifications, repeated here so the deprecation notice makes sense in context: a capable, recent, coding-focused model losing one of its free hosted access paths on short notice.
What to check in your code
Search your codebase and environment for two strings: the base URL integrate.api.nvidia.com/v1 and the model name z-ai/glm-5.3. If both appear together, the banner is addressed to you. The practical questions are the boring ones: which of your calls use that endpoint, what they do, and which alternative model you have actually tested rather than merely heard about. The vendor gives two distinct dates - deprecated on the 27th, unsupported after the 28th - and no cutoff hour, so the only safe reading is to finish the switch before October 27 rather than gamble on what "unsupported" means in practice. If your integration uses Z.ai's own API or a different host, nothing in this notice applies to it; verify which door you use before changing anything.