The short answer
For local weights, use Z.ai’s official Hugging Face repositories: GLM-5 (BF16) or GLM-5-FP8. Z.ai lists both as 744B-A40B models; the FP8 repository lists 756 GB of files. Check the selected repository’s current files and license before downloading. Z.ai’s announcement says the model weights are released under the MIT License. (Z.ai’s model download list, GLM-5 announcement)
If you prefer hosted inference, Z.ai’s announcement says GLM-5 is available through its developer platform. The documented API setup starts with a Standard API key and uses Z.ai’s API endpoint. However, the current generic API example retrieved here uses glm-5.3, not GLM-5; confirm the exact GLM-5 model ID in the live model catalog before sending production requests. (API introduction, GLM-5 API overview)
Official GLM-5 download options
| Repository | Listed variant | What to check |
|---|---|---|
Z.ai GLM-5 |
GLM-5, 744B-A40B, BF16 | Review the model card, files, and license for the revision you plan to use. |
Z.ai GLM-5-FP8 |
GLM-5-FP8, 744B-A40B, FP8 | The Hugging Face repository lists 756 GB of files. Check current repository size and available disk space before starting. |
Z.ai’s table distinguishes BF16 and FP8 but does not, by itself, establish a universal winner for quality, speed, or hardware suitability. Pick the variant supported by your intended serving stack, and check that stack’s instructions before committing to a large download. (Z.ai’s model download list)
Check local requirements before downloading
One documented vLLM route for GLM-5 is substantial: the vLLM recipe specifies vLLM 0.24.0 or newer with CUDA and describes a two-node, 16-GPU tensor-parallel deployment. It lists NVIDIA H200 at 141 GB per GPU and says the multi-node setup requires InfiniBand. The recipe also provides the serving command and per-node instructions; follow those instructions for the selected deployment rather than treating this as a single-machine setup. (GLM-5 vLLM recipe)
For storage, Hugging Face lists 756 GB of files for GLM-5-FP8. The vLLM recipe separately lists an FP8 variant figure of 893 GB. Those are figures from different pages and contexts; do not treat either as a guaranteed free-disk or runtime-memory requirement. Check the current repository size and the full vLLM recipe for your chosen setup. The sources cited here do not specify a separate host-RAM minimum. (FP8 repository files, GLM-5 vLLM recipe)
Download steps
- Choose the official GLM-5 BF16 repository or GLM-5-FP8 repository.
- Review the model card, license, and files for the revision you intend to use. Use the repository’s download controls or the CLI instructions shown there, and download the complete files required by the model.
- Before downloading, compare the available storage with the repository’s current file total. For FP8, the listed Hugging Face total is 756 GB; allow for additional space required by your download workflow.
- Before deployment, compare your environment with the linked vLLM GLM-5 recipe: it specifies vLLM 0.24.0+, CUDA, a two-node 16-GPU setup, H200 hardware, and InfiniBand for multi-node communication. The cited recipe does not give a separate host-RAM minimum.
Download weights or use a hosted API?
Download weights if you need to serve GLM-5 in your own environment and your team can meet the requirements for a documented serving configuration. The vLLM recipe is a specific multi-node reference route, not evidence that a smaller machine will work. (GLM-5 vLLM recipe)
Consider Z.ai’s hosted API if you want managed inference rather than downloading and serving the weights. Z.ai’s GLM-5 announcement says the model is available on its developer platform. To get started, register or sign in to the Z.AI Open Platform, create a Standard API key, and consult the API introduction for the endpoint and Bearer authentication instructions. (Z.ai announcement, API introduction)
Before integrating GLM-5, verify the exact request model ID in the live model catalog or API documentation. The generic chat-completion example currently shown in the API introduction uses glm-5.3; do not assume that is the identifier for original GLM-5. Check current access, limits, and pricing in Z.ai’s documentation before production use. (Chat Completion API)
Availability and source cautions
Ollama’s GLM-5 listing says its entry was retired on July 15, 2026, and that no models had been pushed there. That notice is specific to Ollama; it does not establish availability on Hugging Face or through another provider. (Ollama GLM-5 listing)
The captured SourceForge project page describes GLM-5 but does not show a downloadable artifact or download instructions. Use Z.ai’s upstream repositories for the weights rather than treating that page as a verified download route.
Bottom line: use Z.ai’s Hugging Face links for the local weights, and check the large storage and deployment requirements before downloading. For hosted inference, follow Z.ai’s Standard API-key setup and confirm that the live catalog exposes original GLM-5 under the model ID you intend to call.
Sources
GLM-5 · zai-org GitHub repository
zai-org/GLM-5-FP8 · Hugging Face
GLM-5 Overview · Z.AI Developer Documentation
API Introduction · Z.AI Developer Documentation
Chat Completion · Z.AI Developer Documentation