Modelpublic

0xSero/GLM-5.3-488B-EXL3-TR3-3.42bpw

sourceHugging Faceotherupdated 19d agoView on Hugging Face

1likes336downloads

Model Card

[!TIP] [Support this work →](https://donate.sybilsolutions.ai) · X · GitHub · REAP paper · Cerebras REAP

GLM-5.3-488B — REAP keep-164 (EXL3 TR3 3.42bpw)

GLM-5.3 with 36% of its experts removed — 242 GB, fits within ~300 GB of VRAM. A REAP prune applied directly to davidsyoung/GLM-5.3-EXL3-TR3-3.42bpw, the standard-compliant EXL3 quant (routed experts quantized, sensitive layers BF16).

What this is

GLM-5.3 is a 753B mixture-of-experts model (8 of 256 experts per token, ~40B active). This checkpoint keeps 164 of 256 routed experts per layer, chosen by the max-over-domain saliency criterion (each expert scored by its largest share of any single domain's routed work, so every domain keeps its specialists). Whole experts are dropped, the survivors renumbered, and the routers + MTP layer sliced to match — applied byte-exact to the already-quantized checkpoint, so nothing is re-quantized.


Base	davidsyoung/GLM-5.3-EXL3-TR3-3.42bpw (EXL3 trellis, experts-only quant, BF16 sensitive)
Prune	REAP keep-164/256 (36% removed), max-over-domain saliency, all MoE layers + MTP
Size	242 GB (fits within 300 GB VRAM)
KL vs BF16	measured on GPU — see the fidelity study

Loads in exllamav3 / TabbyAPI, tensor-split across your GPUs.

Credits

—[davidsyoung](https://huggingface.co/davidsyoung) for the EXL3 TR3 3.42bpw base quant this prune is built on.
—[Z.AI / zai-org](https://huggingface.co/zai-org) for GLM-5.3, the base model.
—[Cerebras Research](https://github.com/CerebrasResearch/reap) for REAP (arXiv:2510.13999).
—[turboderp / exllamav3](https://github.com/turboderp-org/exllamav3) for the EXL3 format.

License

Inherits the GLM-5.3 license.