license: mit
base_model: mrfakename/GLM-5.3-Flash-abliterated
pipeline_tag: image-text-to-text
tags:
- gguf
- abliterated
- uncensored
GLM-5.3-Flash (abliterated) GGUF
Huge thanks to Hugging Face for generously sponsoring the compute used to abliterate these models! This was done on HF Jobs, which makes it incredibly easy to run tasks such as these :)
Join the Discord: https://discord.com/invite/foss

GGUF quants of mrfakename/GLM-5.3-Flash-abliterated, which is zai-org/GLM-5.3-Flash with its refusal behaviour removed by directional ablation.
These are the unsloth/GLM-5.3-Flash-GGUF quants with the edit applied in place. Only the edited tensors were dequantized, edited and requantized to their original type. Everything else is byte-identical.
| Folder | Size | Files | Notes |
|---|---|---|---|
Q8_0 |
318 GiB | 8 | Closest to the original weights. |
UD-Q4_K_XL |
186 GiB | 6 | Good default. |
UD-Q2_K_XL |
101 GiB | 4 | Smallest. The edit still survives in full (see below). |
Each folder holds a split GGUF. Point llama.cpp at the -00001-of- file and it loads the rest.
What changed
Chat mode and think mode refuse along slightly different directions, so instead of a single refusal direction a rank-2 refusal subspace was removed:
- Refusal directions were measured separately for chat and think prompts at every layer, on the Q8_0 GGUF.
- The subspace is the top two singular vectors of all of those directions together.
- It was projected out of the residual-stream writes of the 45 main layers:
self_attn.o_proj,mlp.shared_experts.down_proj, andmlp.down_projon the three dense layers. That is 90 tensors, and each layer got its own strength.
The per-layer strengths came from a multi-objective search: minimise refusals and keep the first-token KL on harmless prompts low.
Nothing else was touched:
- Routed experts, embeddings, the vision projector and the MTP/nextn layer are byte-identical to the original quants.
Results (Q8_0 GGUF, greedy)
| refusals, chat (/100) | think mode, reasoning or answer refuses (/100) | KL, harmless first token | |
|---|---|---|---|
| original | 99 | 98 | 0 |
| abliterated | 3 | 2 | 0.038 |
The 100 prompts are held-out harmful requests. Chat replies were 384 tokens; think mode got 768 tokens. Where the reasoning had not finished by then, the reasoning text itself was judged. The KL figure includes a small noise floor from batched inference.
Does the edit survive quantization?
Yes, in every folder here. Low-bit quants can round a small edit away, but these unsloth UD quants keep the edited tensors at Q5_K or better even in UD-Q2_K_XL, and the patch log shows essentially none of the edit lost (at most 2%). See patch-stats.txt.
Running
Mainline llama.cpp runs these; they were patched and tested at commit 2cbfc3c.
Files
Q8_0/,UD-Q4_K_XL/,UD-Q2_K_XL/: the patched quants, e.g.llama-server -m UD-Q4_K_XL/GLM-5.3-Flash-UD-Q4_K_XL-00001-of-….gguf.mmproj-*.gguf: the unchanged vision projector.abliteration-lora.gguf: the edit as a LoRA, if you'd rather use another quant of the original:--lora abliteration-lora.gguf.abliteration-edit.npz: the subspace basis and per-layer strengths.patch-stats.txt: per-tensor patch log.
This model will comply with harmful requests. You are responsible for how you use it.