Mia-AiLab/GLM-5.3-Flash-EXL3-TR3-4bpw
GLM-5.3-Flash TR3 4bpw — mirror (2x DGX Spark recipe)
This is not an original quantization. It is a byte-identical redistribution of brandonmusic/GLM-5.3-Flash-tr3-4bpw at Hugging Face revision `5ab363a8dcf6405955fd5f99671e01a1c9fb124b` so that the 2× DGX Spark recipe MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks stays fetchable if the upstream Hub id moves.
Brandon M. Music created the EXL3/TR3 checkpoint (ShapleyMcg). Z.AI created the base model. Mia's AI Lab only re-hosts this snapshot and serves it on GB10.
@misc{music2026shapleymcg,
author = {Music, Brandon M.},
title = {ShapleyMCG: An Auditable Calibration-to-Encoding Pipeline for
Low-Bit Mixture-of-Experts Models},
year = {2026},
url = {https://github.com/brandonmmusic-max/shapleymcg},
note = {Licensed under the ShapleyMcg License v1.0}
}License
- This checkpoint / ShapleyMcg work: ShapleyMCG License 1.0 (source-available; not OSI “open source”). The
LICENSEfile is the upstream file, unmodified. - Base model zai-org/GLM-5.3-Flash: MIT, Copyright (c) 2026 Z.AI Co., Ltd. Keep that notice with any copy of the base work.
- Spark serve scripts (separate GitHub repo): MIT.
Do not relicense these shards as MIT. Downstream copies of this Work stay under ShapleyMCG License 1.0.
What this is
Uniform-K4 EXL3/TR3 routed-expert checkpoint of GLM-5.3-Flash (~164 GiB, 120 safetensor shards). Provenance metadata from the upstream snapshot is kept.
This Hub repo is not Brandon’s SM120 B12X / NVFP4-KV / EP2/DCP2 daily driver. Do not use the verdictai/glm53-flash-exl3-k4:… image with a “this is the Spark recipe” assumption. His original model card is saved as `ORIGINAL_MODEL_CARD.md`.
Spark recipe (2× NVIDIA GB10, TP=2, fp8 MLA KV, DFlash2):
- GitHub: https://github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks
- Image:
ghcr.io/miaai-lab/glm-5.3-flash-2x-dgx-sparks:exl3
DFlash2 (incoai/GLM-5.3-Flash-DFlash2) is a separate checkpoint under CC BY-NC-ND 4.0 and is not mirrored here.
