Eligible enterprise customers can access the 753-billion-parameter model through managed APIs, with prompt caching and cross-Region inference.
AWS has made Z.ai’s GLM 5.3 available in Amazon Bedrock for eligible enterprise customers. The 753-billion-parameter mixture-of-experts model targets coding and long-running agentic workloads, and Bedrock provides access through managed APIs without requiring customers to operate inference infrastructure.
AWS describes GLM 5.3 as supporting multi-step reasoning, tool-assisted workflows and sustained context across large codebases. Z.ai reports a 50% improvement over GLM 5.2 on its internal coding benchmark and an 84.5 score on the CyberGym security benchmark. These are vendor-reported results. AWS says no direct comparison with GLM 5 was published because the benchmark tests had been updated.
Developers can call the model through OpenAI-compatible Responses and Chat Completions APIs or through Bedrock’s Invoke and Converse interfaces. AWS recommends the OpenAI-compatible options for new applications because they support a more complete feature set. Bedrock can generate API keys for compatible integrations, although AWS recommends short-lived credentials where possible.
Prompt caching is available in two forms. Implicit caching is enabled by default, while explicit caching lets developers mark reusable prompt prefixes such as system instructions, tool definitions or repository files. Each explicit cache breakpoint must cover at least 1,024 tokens. AWS says reusing this stable context can reduce input costs and response latency.
Bedrock offers US and Global cross-Region inference profiles, identified as us.zai.glm-5.3 and global.zai.glm-5.3. Requests originate from a customer-selected AWS Region and are routed by Bedrock for processing. Customers can choose Standard service, lower-cost Flex for less time-sensitive work, or higher-priced Priority service for latency-critical requests.
AWS also demonstrates an authorized security assessment using GLM 5.3 with the open-source Strix agent. The walkthrough targets a locally hosted instance of OWASP Juice Shop, an intentionally vulnerable application. AWS warns users to test only applications they own or have explicit written permission to assess. It also notes that LiteLLM did not yet resolve the global GLM 5.3 identifier at publication, requiring the Converse API route and inference-profile ARN to be specified explicitly.
For existing Bedrock users, the release provides a managed route to a large open-weight model while retaining AWS identity controls, request routing and usage accounting. Production adoption still requires organization-specific validation: eligibility is restricted, service tiers differ in price and latency, and Z.ai’s published benchmark figures do not substitute for testing on a team’s own repositories and workflows.
| Capability | Practical purpose |
|---|---|
| Responses and Chat Completions | OpenAI-compatible application integration |
| Invoke and Converse | Native Amazon Bedrock APIs |
| Prompt caching | Reuse of stable context; explicit segments require at least 1,024 tokens |
| Cross-Region inference | us.zai.glm-5.3 and global.zai.glm-5.3 profiles |
Sources
Originally published by AI Feed: AWS Adds Z.ai’s GLM 5.3 to Amazon Bedrock.
Top comments (0)