GLM-5.3-Flash-NVFP4-Optimized API & Inference Endpoint

What changed

NVFP4 routed-expert weights with repaired shared W1/W3 scale handling.
Reconstruction-optimized NVFP4 scaling.
Selected dense, shared-expert, and attention weights retained in FP8.
Tensors unnecessarily converted to BF16 in the earlier export were restored to their native checkpoint representation.
Tokenizer, processor, chat template, and generation configuration are included.

These changes improve weight reconstruction error relative to the starting NVFP4 export.

Quantization methodology

We started with the Red Hat NVFP4 export and used Z.ai's FP8 checkpoint as the reference. For each routed-expert group, we tested nearby valid NVFP4 scales and kept the one with the lowest reconstruction error. On a representative sample of 72 matrices, mean relative weight RMSE improved from 9.1114% to 7.7185%.

We also restored passthrough tensors that had been expanded to BF16 back to their source FP8 values. W1 and W3 keep shared scales because Marlin serves them as one fused projection; the extra per-matrix scale tuning is applied to W2 only.

Serving

The tested two-node DGX Spark/SM121 recipe is available here:

https://github.com/coolbho3k/GLM-5.3-Flash-NVFP4-DFlash2-2x-DGX-Spark/tree/feature/exl3-ab-current

The recipe supports selectable NVFP4 or EXL3 weights, FP8 or native-FP4 KV cache, DCP2, and optional DFlash2 speculative decoding.

Limitations

The published serving results are specific to the repository's patched vLLM stack on two DGX Sparks.
Independent downstream benchmark evaluation is still recommended.
This is a community derivative and is not an official Z.ai or Red Hat release.

Acknowledgements

Thanks to Z.ai for GLM-5.3-Flash and Red Hat AI for the original NVFP4 checkpoint and quantization work.

What changed

NVFP4 routed-expert weights with repaired shared W1/W3 scale handling.
Reconstruction-optimized NVFP4 scaling.
Selected dense, shared-expert, and attention weights retained in FP8.
Tensors unnecessarily converted to BF16 in the earlier export were restored to their native checkpoint representation.
Tokenizer, processor, chat template, and generation configuration are included.

These changes improve weight reconstruction error relative to the starting NVFP4 export.

Quantization methodology

Serving

The tested two-node DGX Spark/SM121 recipe is available here:

https://github.com/coolbho3k/GLM-5.3-Flash-NVFP4-DFlash2-2x-DGX-Spark/tree/feature/exl3-ab-current

The recipe supports selectable NVFP4 or EXL3 weights, FP8 or native-FP4 KV cache, DCP2, and optional DFlash2 speculative decoding.

Limitations

The published serving results are specific to the repository's patched vLLM stack on two DGX Sparks.
Independent downstream benchmark evaluation is still recommended.
This is a community derivative and is not an official Z.ai or Red Hat release.

Acknowledgements

Thanks to Z.ai for GLM-5.3-Flash and Red Hat AI for the original NVFP4 checkpoint and quantization work.

GLM-5.3-Flash-NVFP4-Optimized

README

What changed

Quantization methodology

Serving

Limitations

Acknowledgements

Explore FriendliAI today

README

What changed

Quantization methodology

Serving

Limitations

Acknowledgements