GLM 5.3 lands on Amazon Bedrock with full 1M context
AWS adds Z.ai's top open coder to Bedrock with IAM, guardrails and prompt caching—but it's cross-region and invite-only. Specs, the catch, a self-host path.
Copy markdown
Call it with your AWS keys, no Z.ai account
GLM 5.3 answers on Bedrock as `zai.glm-5.3` through the Converse, Invoke, Responses and OpenAI-compatible Chat Completions APIs, so it inherits IAM auth, VPC, guardrails, agents and prompt caching. For AWS shops, a near-frontier open coder drops into your existing stack with no separate Z.ai contract or billing.
The catch: cross-region only, and invite-gated
There's no in-region on-demand access—5.3 runs exclusively via cross-region inference (`us.zai.glm-5.3` or `global.zai.glm-5.3`), and it's limited to “eligible customers” who must ask their AWS account team. Your prompts can leave the home region, so check data-residency rules before wiring it into production.
Full 1M context—double what GLM 5 got on Bedrock
Bedrock serves the complete 1M-token window plus 128K output, where the older GLM 5 was capped at 200K there. That's whole-repo review and long agent traces inside AWS, available on Standard, Priority and Flex tiers (Reserved isn't supported).
What you're getting: a post-training coding leap
5.3 reuses GLM-5.2's base—every gain is post-training. Z.ai reports Terminal-Bench 3.0 jumping 4.6→28.3, DeepSWE v1.1 46.2→66.9, and ExploitBench 24.4→54.4%. It's a 744B-parameter MoE firing ~40B active per token, tuned for long-horizon agent work.
Reasoning defaults to 'max'—mind the token bill
The Invoke body takes `reasoning_effort`, and 5.3 defaults to max with no way to fully switch thinking off. Bedrock's own rates aren't published yet; on third-party gateways 5.3 runs about $1.40 in / $4.40 out per 1M tokens (ZenMux), so price the reasoning overhead before you route real traffic.
Want to self-host instead? Reach for 5.3 Flash
The flagship weights stay API-only, but GLM 5.3 Flash (320B MoE, MIT license) is on Hugging Face with day-one vLLM and SGLang support, around $0.15 in / $0.50 out per 1M on Fireworks. That's the self-hostable path if the Bedrock invite wall or residency rules block you.