Z.ai's GLM 5.3 goes GA on Amazon Bedrock for eligible enterprise customers, via cross-Region inference only

Why it matters

【Key points】

AWS says GLM 5.3 from Z.ai (Zhipu AI) is generally available on Amazon Bedrock, posted October 5, 2026, but only "to eligible enterprise customers"; the model card tells others to contact their AWS account team. Z.ai confirmed it on X on October 6, Beijing time.

The model offers a 1M-token context and up to 128K output tokens, with selectable reasoning effort and prompt caching. It runs only through cross-Region inference, via the US (us.zai.glm-5.3) or Global (global.zai.glm-5.3) profile; in-Region inference is not supported.

【Verifiable】

Primary: https://aws.amazon.com/about-aws/whats-new/2026/10/amazon-bedrock-glm-5-3/ , the AWS launch blog and the Bedrock model card (model-card-zai-glm-5-3).

AWS gives two parameter counts. The announcement says "753B total parameters and roughly 40B active per token", and the blog also says 753B; the model card says "744B total parameters and approximately 40B active per token". Z.ai's Hugging Face card states no total.

The card lists explicit caching from 1,024 tokens per checkpoint and US Geo inference from four US Regions.

The card links its licence to the Apache 2.0 file in Z.ai's GLM-5 repository, but the GLM-5.3 weights carry a GLM-5.3 License: a model-as-a-service operator with group revenue above $10 billion over 12 months must pass Z.ai's security review before commercial use.

【Our read】

A US hyperscaler's managed channel gives a Chinese lab distribution its own API cannot reach: enterprises call the model under their AWS controls instead of Z.ai's own API. It also puts access in AWS's hands.

The eligibility gate and US/Global-only routing read as a cautious rollout. AWS does not define eligibility, and customers needing EU or APAC data residency have no Geo option.

The 753B/744B split is minor; Z.ai's own card gives no total, so confirm with Z.ai, not a cloud listing.

Watch whether access widens, whether EU or in-Region options arrive, and how the security-review clause applies to clouds.