# GLM-5.3: Self-Hosting Hardware, Scenarios and Commercial License

> Frontier text-only coding, cybersecurity, tool-use and long-horizon agent workloads that can support an eight-accelerator FP8 worker or a multi-node BF16 deployment.

- Verified: 2026-09-12
- Released: 2026-08-28
- Canonical: https://chinaapi.ai/open-model-deployment/glm-5.3/
- Back to directory: https://chinaapi.ai/open-model-deployment/
- Machine-readable dataset: https://chinaapi.ai/data/open-model-deployment.json

## Model facts

- Vendor: Z.ai
- Parameters: 753B checkpoint; official serving guides describe an approximately 743B-744B architecture total; approximately 39B-40B per token active
- Native precision: The default official checkpoint is native FP8. A separate official BF16 variant needs about 1.5 TB for weights and is a multi-node serving target outside the largest current single-node systems.
- Context: 1M tokens
- Category: General and agentic foundation models
- Modalities: text → text
- Serving paths: SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth, vLLM-Ascend, xLLM
- Official weights: Official FP8 / default: https://huggingface.co/zai-org/GLM-5.3?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment; Official BF16: https://huggingface.co/zai-org/GLM-5.3-BF16?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment
- Model documentation: https://github.com/zai-org/GLM-5?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment

## Deployment evidence

- **Official launch / minimum — evidence B:** vLLM documents the native FP8 checkpoint on one 8x H200 or 8x H20 node. That is the smallest current official vLLM topology in this dossier, not an absolute minimum or a production capacity claim. Source: https://recipes.vllm.ai/zai-org/GLM-5.3?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment
- **200-person team — evidence B:** For the defined 200-seat profile, start load testing one 8x H200 FP8 worker; use 8x B200 when the full 1M context is part of the target mix. Add a second independent worker only when maintenance or failure recovery is required. No official recipe maps employee count to capacity. Source: https://recipes.vllm.ai/zai-org/GLM-5.3?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment
- **Commercial API — evidence B:** Use independently deployable FP8 workers with admission control and separate ordinary-, long-context- and reasoning-effort SLOs. The vLLM recipe uses 8x H200/H20 for standard serving and 8x B200 with FP8 KV cache for a full-1M starting point; BF16 requires a multi-node design. Source: https://recipes.vllm.ai/zai-org/GLM-5.3?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment

## Scenario evidence

- **Complex coding and long-horizon agents — Vendor-stated:** Z.ai positions GLM-5.3 as a post-trained GLM-5.2 derivative for complex coding and long-horizon agent work. The reported improvements are vendor evaluations, not ChinaAPI reproductions. Source: https://huggingface.co/zai-org/GLM-5.3?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment
- **Cybersecurity analysis — Public evidence:** The official card publishes CyberGym, ExploitGym and ExploitBench results with named harnesses and evaluation settings. They support candidate selection but do not establish production safety or local quality parity. Source: https://huggingface.co/zai-org/GLM-5.3?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment
- **Private million-token codebase analysis — ChinaAPI inference:** The text-only 1M context and official self-hosting paths make it a candidate for codebases and records that must remain inside the operator's boundary; practical batch size still depends on KV-cache and latency measurements.

## Commercial-use check

- License: GLM-5.3 License — https://huggingface.co/zai-org/GLM-5.3/blob/main/LICENSE?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment
- MaaS / hosted service: Commercial use is granted, but a licensee or affiliate that operates a Model-as-a-Service business and exceeds USD 10 billion in aggregate revenue over any consecutive 12-month period must pass Z.AI's security review before commercial use.
- Attribution: Retain the copyright and license notice in copies or substantial portions of the weights, comply with applicable laws and the license's use restrictions, and note that the weights are provided without warranty.

## Avoid or validate first

- Treating the roughly 39B-40B active-parameter figure as the weight-memory footprint; the full expert checkpoint must be loaded
- Attempting the official BF16 checkpoint on a single 8x H200 node
- Operating a qualifying MaaS business above the license's USD 10 billion threshold without the required Z.AI security review
- Leaving maximum reasoning effort enabled for all latency-sensitive traffic without measuring output length and service cost

## Known limitations and open questions

- No ChinaAPI hardware reproduction, quality benchmark or like-for-like comparison between the downloadable checkpoint and hosted route
- The full-1M vLLM example on 8x B200 starts at max-num-seqs 32 and explicitly requires tuning to available KV-cache headroom; it is not a concurrency guarantee
- The 753B checkpoint count and the approximately 743B-744B architecture shorthand come from different official reporting layers and should not be treated as a memory estimate
- Thinking is always enabled and defaults to maximum effort unless low or high is passed explicitly
- Vendor benchmark and capability claims have not been reproduced by ChinaAPI

## Evidence boundary

Loading weights, completing a first forward pass and meeting a production latency/SLA are separate thresholds. The 200-person and commercial tiers require workload-specific measurement. License summaries are product research, not legal advice.

Generated by GENERATED BY scripts/gen_open_model_deployment.py.