GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
soyrsoyr
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
soyrsoyr
Model Tree
zai-org/GLM-5.3
Input Modalities
Output Modalities
Supported Functionality
Full pretrained checkpoint converted from zai-org/GLM-5.3 using llm-compressor PR3118, 572c6390.
Checkpoint structure and MTP tensor checks passed. Runtime validation and acceptance-rate measurement are pending. No quality or performance benchmark is claimed.
The backbone MLP gate/up/down projections and supported MTP attention and MLP projections use data-free NVFP4A16: weight-only FP4 with 16-bit activations. Other projections remain BF16. Native FP8 source weights are dequantized before conversion. The backbone uses model_free_ptq with two GPUs, followed by PR3118’s prepare_mtp_save / save_mtp_tensors integration for MTP. The model reference contains only meta tensors; this run does not test a full-model oneshot load. This run replaces MFPTQ’s legacy per-shard fusion diagnostic with complete-index partner and module-identity checks; each shard worker also explicitly sets its assigned CUDA device for Triton packing. Quantization math and dependency resolution are unchanged. See config.json and pr3118-validation.json for the configuration and checks.
The output generation defaults enable sampling to make the source top_p=0.95 setting valid under Transformers 5.17.
The source model's licensing applies; consult its model card and included license files.
This artifact is intended for PR validation. It can be downloaded later for runtime testing.