# GLM 5.3 arrives on Amazon Bedrock - Source: MechNext — https://mechnext.eu/en/article/31-glm-5-3-arrives-on-amazon-bedrock - Published: 2026-10-06 - Updated: 2026-10-06 - Category: Industry News - Language: en (Markdown in other languages: ?lang=de, ?lang=sk, ?lang=pl, ?lang=zh, ?lang=ru) > AWS says GLM 5.3 from Z.ai is available to eligible Amazon Bedrock enterprise customers. Its announcement outlines the model, API options, prompt caching, service tiers, and a security-testing walkthrough. The AWS Machine Learning Blog article “Introducing GLM 5.3 on Amazon Bedrock” was published on October 5, 2026. AWS announces that GLM 5.3 from Z.ai, also known as Zhipu AI, is available on Amazon Bedrock to eligible enterprise customers. The announcement concerns access through that service for that customer group. AWS describes GLM 5.3 as a 753-billion-parameter mixture-of-experts model optimized for coding and long-horizon agentic tasks. Those two areas frame the introduction: working with code and carrying out agentic tasks that extend across multiple steps. The model size and stated focus are characteristics reported in the article. They do not, by themselves, establish how the model will perform in every particular application. ### Ways to invoke the model The article says GLM 5.3 can be invoked through the OpenAI-compatible Responses and Chat Completions APIs, as well as the Amazon Bedrock Invoke and Converse APIs. These are the invocation paths named in the announcement. For an application team, having those options means the integration can be considered in relation to the interfaces it already uses. The article identifies the APIs; choosing among them depends on the design and needs of the application. The announcement also describes prompt caching. Implicit prompt caching is supported by default, while explicit cache controls are available on the Responses and Chat Completions APIs. This distinguishes the default support from the direct controls the article associates with those two APIs. Caching is therefore part of the described access experience, alongside the invocation interfaces. The announcement does not prescribe a cache configuration for every workload, so teams must consider how caching fits their own use of prompts. ### Inference profiles and service tiers For GLM 5.3, the article lists a US cross-Region inference profile, `us.zai.glm-5.3`, and a Global cross-Region inference profile, `global.zai.glm-5.3`. These identifiers make the two named profile options explicit. The article connects them to GLM 5.3 but does not provide additional operating characteristics for each profile in the supplied information. AWS describes three service tiers. Flex is for workloads that are less time-sensitive. Priority is for latency-critical requests and comes at a higher price. Standard is the default, balancing price and speed. The descriptions give users a straightforward way to think about the choice: a workload’s sensitivity to time and its price-versus-speed balance are relevant considerations. The announcement describes the tiers’ intended roles rather than identifying one tier as right for all workloads. These tier definitions are useful when relating an API integration to operational priorities. A task that is not time-sensitive aligns with the stated purpose of Flex; a latency-critical request aligns with Priority, with its higher price; and Standard is the stated default balance. The article supplies these distinctions but no workload-specific recommendation or numerical performance comparison among the tiers. ### Reported measurements and the Strix walkthrough At release, Z.ai measured a score of 84.5 on the CyberGym benchmark, according to the article. Z.ai also reported a 50% improvement over GLM 5.2 on its internal coding benchmark. Both results are attributed to Z.ai in the announcement. One is tied to a named benchmark and the other to an internal coding benchmark, so they should be understood as reported measurements rather than independent confirmation of performance in every setting. The article also describes Strix as an open-source AI penetration-testing agent. Its walkthrough uses the OWASP Juice Shop sample application running locally. This example places the model announcement alongside a concrete security-testing scenario: the named agent and a local sample application are part of the walkthrough. That description does not establish outcomes for other targets or applications; it identifies what the article uses to illustrate the scenario. ### What the announcement gives enterprise readers For eligible enterprise customers using Amazon Bedrock, the article brings together several practical pieces of information: the model’s stated focus, four named API options across OpenAI-compatible and Bedrock interfaces, prompt-caching support, two inference profile identifiers, and three service tiers. Together, these details describe ways to access the model and choices that can shape an integration. They let a reader identify which subjects to evaluate when considering GLM 5.3 for coding or longer agentic tasks. The reported scores add context, with attribution that matters. The CyberGym result and the internal coding comparison are presented as Z.ai measurements or reports. They do not amount to a promise about a particular customer workload. Likewise, the Strix walkthrough is described using a local OWASP Juice Shop sample application, which gives the example a defined scope. In summary, AWS presents GLM 5.3 as available on Amazon Bedrock to eligible enterprise customers and characterizes it as a large mixture-of-experts model focused on coding and long-horizon agentic work. The article explains invocation options, default and explicit prompt-caching capabilities, cross-Region profiles, and the purposes of Flex, Priority, and Standard. It also reports Z.ai’s benchmark claims and describes a Strix walkthrough with a locally running sample application. These are the announcement’s core details for readers assessing the service’s stated capabilities and integration choices.