AWS Puts Z.ai’s GLM-5.3 on Bedrock; Zhipu Shares Jump on a Reported Revenue Share

A Chinese frontier lab’s open-weight model is now a managed, U.S.-routable product inside the cloud many U.S. enterprises already buy from. For a buyer wary of calling a Chinese lab’s API directly, inference now runs under its own AWS account and controls, which is a different procurement question. If the reported revenue share is real, Bedrock also becomes an overseas revenue line for a listed Chinese lab.

In an October 5 post, AWS said GLM 5.3 from Z.ai (Zhipu AI) is available on Amazon Bedrock to eligible enterprise customers. AWS describes it, as published on Hugging Face, as a 753B-parameter mixture-of-experts model optimised for coding and long-horizon agentic tasks. On October 6, Zhipu (02513.HK) was up 7.7% at HK$716 in morning trade, its second straight day of gains, according to etnet.

What AWS listed

• APIs: OpenAI-compatible Responses and Chat Completions, plus Bedrock’s Invoke and Converse. AWS recommends the OpenAI-compatible APIs for new applications.
• Routing: US (us.zai.glm-5.3) and Global (global.zai.glm-5.3) cross-Region inference profiles.
• Prompt caching: implicit by default, with explicit cache breakpoints on the OpenAI-compatible APIs.
• Service tiers: Flex, Standard and Priority.
• Access: “eligible enterprise customers,” not every Bedrock account. AWS doesn’t say what makes a customer eligible.

AWS’s benchmark lines are Z.ai’s: competitive coding results, a 50% gain over GLM 5.2 on Z.ai’s internal coding benchmark, and 84.5 on CyberGym at release. AWS frames the cyber results as “a natural fit for defensive security workflows.”

The security demo

AWS’s walkthrough runs Strix, an open-source AI penetration-testing agent, with GLM 5.3 against OWASP Juice Shop, a deliberately vulnerable app running on the user’s own machine. The post says to test only applications you own or have written permission to test. That framing matters given what is already known about this model: Anthropic’s red team reported that GLM-5.3 nearly matches Claude Mythos Preview on some exploit-building tests and that its safeguards fall to simple bypasses, and NIST’s CAISI called it the most cyber-capable open-weight model released to date (Times of AI, September 29). AWS is selling authorised testing under a customer’s own AWS controls, not endorsing offensive use. But it has made a cyber-capable open model a one-line API call for enterprise accounts.

The revenue share is reported, not announced

etnet, citing an unnamed Chinese media report, said AWS will share revenue with Zhipu based on model usage, that Zhipu has set up similar arrangements with several other overseas cloud providers, and that at home it has revenue-share deals with platforms including Alibaba Cloud’s Bailian, while Huawei Cloud has listed GLM-5.3. None of that is in AWS’s post, and neither company has confirmed terms. The share move happened, but what drove it is a reported deal.

Limits

• The revenue-share terms come from etnet’s summary of an unnamed Chinese media report. AWS’s post doesn’t mention them.
• Benchmark figures in AWS’s post are Z.ai’s own, not Bedrock-specific or independently run.
• AWS doesn’t publish pricing in the post or define “eligible enterprise customers.”
• The share price is an intraday morning figure, not a close.

Sources

• AWS Machine Learning Blog: Introducing GLM 5.3 on Amazon Bedrock (October 5, 2026)
• etnet: Zhipu surges nearly 8%, reportedly integrates GLM-5.3 large model into Amazon cloud platform to pursue overseas revenue (October 6, 2026)
• Times of AI: Anthropic red team on GLM-5.3’s exploit-building and safeguards (September 29, 2026)