GLM 5.2 FAQ

FP8 requires ~372 GB VRAM — you need 4×H100 (80GB each) minimum. 8×H200 is the recommended production setup. With Q4_K_M GGUF and CPU offloading, you can run on 4×RTX 4090 (96GB total + 128GB RAM) at 2-5 tok/s — usable for testing, not production.

GLM 5.2 was trained on Huawei Ascend, but inference today works on NVIDIA GPUs via vLLM and SGLang. Ascend inference support is in development but not production-ready. AMD ROCm support is not yet available for GLM 5.2's custom architecture.

4×H100 (80GB) with FP8 quantization. This gives you ~128K context with 4-8 concurrent requests at 25-50 tok/s per user. For 1M context at production throughput, 8×H200 is the realistic minimum.

FP8 weights are ~744 GB. On a 10 Gbps connection: ~40 minutes. On 1 Gbps: ~2 hours. The model is pulled from HuggingFace Hub on first vLLM start — use huggingface-cli download to pre-cache for faster restarts.

Yes. Full MIT license — no use-based restrictions, no downstream propagation requirements. Unlike DeepSeek's custom license (which has Paragraph 5 use-based restrictions flagged by Black Duck), GLM 5.2's MIT is the most permissive open-source license available. Deploy it, embed it, redistribute it — no legal review needed.

Yes. MIT license permits derivative works for commercial use without restriction. Note: fine-tuning a 744B model is a significant undertaking — this guide covers inference and serving, not training.

vLLM: broader ecosystem, Day-0 GLM 5.2 support, more community examples. SGLang: RadixAttention provides 15-30% better throughput for long-context workloads via mixed chunk prefill. If your workload is >32K context, try SGLang. Both are production-ready.

Similar VRAM requirements (both ~370GB FP8). GLM 5.2 advantages: MIT license (no legal review), 1M native context, #1 Code Arena. DeepSeek advantage: larger ecosystem, more community deployment examples. If commercial deployment without license review matters, GLM 5.2 is the clear choice.

Z.ai API: $30-80/month (Pro/Max plans). Self-hosting: 8×H200 rental ~$25-35/hr on cloud. For continuous serving, self-hosting is more expensive unless you own hardware. For bursty CI/CD workloads (code security pipeline), self-hosting can be cheaper because you only pay when running. The manual includes a break-even calculator.

The model itself runs on your infrastructure — no code leaves your network. The security concern is prompt injection: a malicious file could attempt to manipulate GLM 5.2's output. The manual covers input sanitization, sandboxed runners, and prompt engineering for security audit that minimizes injection risk.

No. GLM 5.2 was open-sourced June 16, 2026 — 14 days ago as of this writing. No teams have publicly documented production deployments. This guide is built from first-principles testing on 8×H200 and 4×H100. When production case studies emerge, we'll update the guide.