GLM-5.3 Uncensored EXL3
EXL3 conversions of dealignai/GLM-5.3-UNCENSORED-FP8.
| Folder | Weight size | Attention/shared experts | MTP |
|---|---|---|---|
| 3bpw | 307.59 GB | 5-bit protected | 16-bit |
| 6bpw | 585.17 GB | 8-bit protected | 16-bit |
Both use ExLlamaV3's high-quality allocation, so overall bitrate is slightly above the base target. See each folder's README for conversion details. Download a complete quant folder, including all shards and configuration/tokenizer files. Uploads may appear incrementally; ensure every shard listed in the index has downloaded before loading.
MTP is retained inside each quant. Optional DFlash2 is provided in its own folder, copied from incoai/GLM-5.3-DFlash2 at revision 425aa615ce320caac34400208b30808c8f14f76c. It is not a draft model retrained for the uncensored variant. Its original model card is retained. Compatible inference software and sufficient RAM/VRAM are required.
These checkpoints derive from an FP8 source. Consult the source and upstream GLM-5.3 repositories for applicable licensing and model information.