The Pulse
AWS Brings Z.ai’s 753B-Parameter GLM 5.3 to Bedrock
AWS added Z.ai’s GLM 5.3 to Amazon Bedrock on October 5, 2026, offering the model through managed APIs and cross-Region inference. Access is limited to eligible enterprise customers, and AWS highlights coding, agentic workflows, and prompt
AI.info Team · Written with AI ·
At 753 billion parameters, Z.ai’s GLM 5.3 is now available through Amazon Bedrock, where AWS says customers can run the mixture-of-experts model without managing inference infrastructure. AWS announced the addition on October 5, describing the model as geared toward coding and long-running agentic tasks. Access is currently restricted to eligible enterprise customers.
Managed access, with an enterprise gate
GLM 5.3 can be called through Amazon Bedrock’s Invoke and Converse APIs, as well as OpenAI-compatible Responses and Chat Completions APIs. AWS recommends the OpenAI-compatible options for new applications, saying they provide a more complete set of features. Developers can also try the model in the Bedrock console’s Playground.
“Access to GLM 5.3 on Bedrock is available to eligible enterprise customers,” AWS AI Specialist Solutions Architect Alex Thewsey writes in the announcement.
The service offers US and global cross-Region inference profiles, which let Bedrock route requests for processing across AWS Regions. The US profile is identified as us.zai.glm-5.3; the global profile is global.zai.glm-5.3. Customers can choose among Standard, Flex and Priority service tiers, trading off price and speed according to workload needs.
Built for repeated context and long tasks
AWS foregrounds coding and agent workflows that carry substantial context between model calls, such as repository-wide changes or multi-step tasks that use tools. GLM 5.3 supports implicit prompt caching by default. Developers can also mark reusable prompt sections explicitly; AWS says each cache breakpoint must contain at least 1,024 tokens to qualify.
That feature targets a recurring expense in agentic applications: resending stable instructions, tool definitions or code context on each turn. AWS says caching can lower input costs and latency when requests reuse an initial prompt prefix, though the announcement does not provide a price comparison or measured latency results.
Security claims come from Z.ai
AWS also points to GLM 5.3’s reported cybersecurity performance. The announcement says Z.ai measured a score of 84.5 on the CyberGym benchmark and describes the model as suitable for defensive security work. AWS attributes the score to Z.ai; it does not present an independent evaluation in the post.
To demonstrate a security workflow, AWS walks through running Strix, an open-source AI penetration-testing agent, against OWASP Juice Shop, a deliberately vulnerable application hosted locally. The walkthrough routes model inference through Bedrock, under the customer’s AWS account controls, and warns users to test only systems they own or have written permission to assess.
What AWS’s launch leaves open
Z.ai’s benchmark claims and AWS’s implementation details answer different questions: the former describe model performance, while the latter lay out how customers can access and operate it on Bedrock. The launch post does not include independent benchmark results, comparative pricing, or a broader availability date beyond eligible enterprise customers.
For now, the concrete offer is a managed endpoint with multiple API paths, prompt caching, cross-Region routing and three service tiers. Enterprise customers can invoke GLM 5.3 from the console or through supported APIs; other customers are not included in AWS’s stated access eligibility.