Models/Blackfrost/BF-GLM53-NVFP4-001

GLM-5.3 — De-Risked NVFP4

Public repositoryFree · Ungated

Z.AI · GLM-5.3 · NVFP4 · free public download

README.mdBF-GLM53-NVFP4-001 · free public access
text-generationnvfp4mixture-of-expertspublicfree

GLM-5.3 — De-Risked NVFP4

The complete 256-expert GLM-5.3 NVFP4 package with a documented 8× NVIDIA B200 target.

Download free ↓
Free and ungated

Download the public weights directly from Hugging Face. No payment, account, or access request is required. Public GLM-5.2 and GLM-5.3 downloads remain available independently of the paid store.

License terms

Requires the Z.AI copyright and permission notice with copies or substantial portions. MaaS operators and affiliates above the stated $10B revenue threshold must complete Z.AI security review before commercial use. The repository’s existing license terms and notices still apply. Read the upstream license ↗

Release identity and selection guidance

Edition roleDeployment quant
Weight package~432.9 GiB · ModelOpt W4A4 NVFP4 · 158 model safetensors shards + 3 MTP shards
Intended fitThe documented SGLang 0.5.17, TP8, 8× NVIDIA B200 serving configuration.
Choose this edition whenChoose this full-topology quant when the documented B200/SGLang target matches your environment and the BF16 footprint is not required.
Deployment boundaryThe purchase supplies the NVFP4 weights and repository artifacts, not SGLang, GPU capacity, a running endpoint, or operating support.
Verified repository revisionec576d33fe18 ↗

Technical specifications

Base modelPrior Blackfrost GLM-5.3 BF16 master
ArchitectureGlmMoeDsaForCausalLM
Release identifierBF-GLM53-NVFP4-001
Parameters~753B topology · 256 routed experts · no pruning
PrecisionModelOpt W4A4 NVFP4 mixed precision
Footprint~432.9 GiB · 158 model shards + 3 MTP shards
ServingSGLang 0.5.17 · 8× NVIDIA B200 · TP8

What is available

The model weights and supporting files published in the named repository, including configuration, tokenizer files when present, and documentation. Review the model card for its current file inventory and usage terms.

Compute, hosting, installation, custom engineering, and ongoing support are separately scoped. Explore hosting or custom services.

Deployment notice

Validate model behavior, runtime compatibility, infrastructure capacity, and operational controls against your workload. Listed configurations are references, not performance guarantees.