GLM-5.3-Flash EXL3 4bpw TensorFold Ablit
EXL3 4bpw TensorFold weights of GLM-5.3-Flash. Layers 15–43 and MTP layer 45 replace self_attn.o_proj with the Keys tensors. Layers 0–14 and layer 44 stay the parent. Experts, vision, QKV, embeddings, and the head are the parent weights.
drowzeys (GitHub) authored those o_proj tensors. They are the files in drowzeys/keys-GLM-5.3-Flash-NVFP4-ablit-l15-43-mtp-l45 at commit cecebc0ff0bbce8eefaa19667290a79d0216cd0e. Method and scripts: drowzeys/keys-GLM-5.3-Flash-NVFP4-ablit-l15-43-mtp-l45.
dealignai (@dealignai) authored GLM-5.3-Flash-UNCENSORED-NVFP4. jordanschenck ran the compute for that checkpoint. drowzeys published the tensors copied here as an altered copy of that checkpoint's BF16 self_attn.o_proj (layers 15–43 and MTP layer 45; layers 0–14 and 44 left stock).
Mia's AI Lab authored the parent quant, Mia-AiLab/GLM-5.3-Flash-EXL3-4bpw-TensorFold at commit 078455ffe6472f9a52fbc1139f58b9db2881b25c.
Z.AI authored GLM-5.3-Flash. turboderp authored exllamav3. Serving engine: TensorFold.
This repository is gated. Agree to the Responsible Use terms on the form above before download. The license is MIT, inherited from Z.AI GLM-5.3-Flash. See RESPONSIBLE_USE.md.
Credits
- drowzeys authored the Keys
o_projtensors copied into this checkpoint: Hugging Face commitcecebc0ff0bbce8eefaa19667290a79d0216cd0e, GitHub, profile huggingface.co/drowzeys and github.com/drowzeys. - dealignai authored GLM-5.3-Flash-UNCENSORED-NVFP4. Profile: huggingface.co/dealignai, @dealignai. jordanschenck ran the compute for that checkpoint: @jordanschenck.
- Mia's AI Lab authored the EXL3 4bpw parent: Mia-AiLab/GLM-5.3-Flash-EXL3-4bpw-TensorFold at
078455ffe6472f9a52fbc1139f58b9db2881b25c. - Base model: Z.AI, GLM-5.3-Flash (MIT).
- Quantization format and converter: exllamav3 by turboderp (MIT).
- Serving engine: TensorFold.
What changed
| Parent | Mia-AiLab/GLM-5.3-Flash-EXL3-4bpw-TensorFold @ 078455ffe6472f9a52fbc1139f58b9db2881b25c |
o_proj source |
drowzeys Keys commit cecebc0ff0bbce8eefaa19667290a79d0216cd0e |
| Edited | BF16 model.language_model.layers.{15–43,45}.self_attn.o_proj.weight (30 tensors) |
| Left stock | Layers 0–14 and layer 44 o_proj; all experts, vision, QKV, embeddings, head |
| Relative change | min 0.0415, median 0.115, max 0.187 (layer 43) |
| Layout | 83 EXL3 shards, same config.json as the parent |
| License | MIT (Z.AI GLM-5.3-Flash) |
drowzeys reports 32/32 Refusal32 bypass, 0 refuse, 0 garble on the NVFP4 parents those tensors were published for, with thinking off and no stock drafter. That score is theirs, measured on those NVFP4 trees. This file is the same 30 tensors copied into the Mia EXL3 tree. It does not add a new Refusal32 number.
ABLIT_META.json records each edited tensor, shard offset, and sha256. Layer 15 sha256 begins 1b46d7388815c0d7. Layer 45 sha256 begins 91f809721dd7f90a.
Use
Serve it with the same GLM-5.3-Flash EXL3 2× DGX Sparks TensorFold recipe as the parent. Point the recipe at this directory. Do not pass a different quantization format.
The stock chat_template.jinja is Z.ai's. It opens <think> even when thinking is turned off. drowzeys authored a closed-think overlay for that behavior: chat_template.thinking-off.jinja.
hf download Mia-AiLab/GLM-5.3-Flash-EXL3-4bpw-TensorFold-Ablit \
--local-dir ~/models/GLM-5.3-Flash-EXL3-4bpw-TensorFold-Ablit
Download stays closed until you submit the gate form on the model page.