license: mit
base_model:
- mrfakename/GLM-5.3-Flash-abliterated
base_model_relation: quantized
pipeline_tag: image-text-to-text
library_name: gguf
tags: - gguf
- ds4
- dwarfstar
- abliterated
- glm
GLM-5.3-Flash-abliterated — DS4 Q2
This conversion is packaged for antirez/ds4 (DwarfStar), including its Metal backend on Apple Silicon Macs. Use a current DS4 build with GLM-5.3 Flash support.
The weights come from mrfakename/GLM-5.3-Flash-abliterated, pinned at 7c10d597d8711d6f3ca65539738447eaa1ba8b89. Credit for the original model goes to Z.ai; credit for abliteration and the original results below goes to mrfakename. This repository provides the DS4 conversion.
Why I chose these source models
Reducing refusals is only part of the job. A model also needs to finish its reasoning and deliver a usable answer. A permissive model can still spend its entire output budget reasoning in circles without producing a final answer. Retaining reasoning, knowledge and coding ability matters too.
In my related DeepSeek V4.1 Flash 1017 comparison, I tested full checkpoints through SGLang on an H200, at temperature 0 with a 16k output-token cap. My follow-up includes MoonMath.
| Test | Original | Mrfakename | Pqhaz heretic | MoonMath |
|---|---|---|---|---|
| HarmBench chat | 16.3% | 93.8% | 93.8% | 92.1% |
| HarmBench thinking | 3.8% | 86.7% | 72.9% | 82.5% |
| MITRE substantive assistance | 16.2% | 84.2% | 82.8% | 83.1% |
| Benign-prompt refusals | Not reported | 0/750 | 0/750 | 0/750 |
HarmBench covered 240 text cases per mode, excluding copyright cases, using its official classifier. These percentages measure benchmark success, not general intelligence. MITRE figures are provisional results from a custom audited rubric judged by the original model, not official CyberSecEval scores; generated code was not executed. Benign-prompt refusals were detected heuristically; the original model's count was not reported.
The reported thinking runs with no final answer were:
| DeepSeek V4.1 Flash 1017 edit | Cases with no final answer |
|---|---|
| mrfakename abliterated | 21 |
| MoonMath uncensored | 30 |
| pqhaz heretic | 45 |
This comparison is why I chose mrfakename's checkpoints for these conversions: fewer unfinished thinking runs in those tests, alongside reduced refusals. It did not eliminate the problem.
Scope: These are related DeepSeek results from full checkpoints, not benchmarks of this GLM or Qwen conversion, its quantization, or DS4. They do not establish a universal ranking of builders or methods. Knowledge and coding retention were not evaluated in that comparison, so these results do not establish that capability loss is absent.

Files and quantization
GLM-5.3-Flash-abliterated-Q2.gguf— approximately 96.51 GB / 89.88 GiB; includes the MTP block.GLM-5.3-Flash-abliterated-Vision-Encoder.gguf— separate BF16 vision encoder, approximately 1.13 GB.SHA256SUMS,build-provenance.json, andconverter.patch— integrity and conversion provenance.
IQ2_XXS gate/up routed experts and Q2_K down routed experts, with mixed higher-precision tensors elsewhere. The original author changed 90 residual-write tensors and left experts, embeddings, vision and MTP unchanged. This conversion regenerates all 90 edited tensors directly from the pinned abliterated checkpoint and copies every other tensor payload from the published DS4 Q2 model, pinned at b2fa29d7a6b410db11221c904973967b80b760f5. All output payloads are checked against either the abliterated source conversion or that reference file. No LoRA application is needed.
This is one complete model file: no concatenation or split-file reconstruction is needed. The Q4 and Q2 versions are published separately.
Run with DS4
git clone https://github.com/antirez/ds4.git
cd ds4
make -j8
hf download daemonix/GLM-5.3-Flash-abliterated-Q2-GGUF-antirez-DS4 --local-dir ./models/glm53-abliterated-q2
./ds4 -m ./models/glm53-abliterated-q2/GLM-5.3-Flash-abliterated-Q2.gguf --ctx 32768
Add --mtp to use the embedded draft block. Add --vision ./models/glm53-abliterated-q2/GLM-5.3-Flash-abliterated-Vision-Encoder.gguf to enable image input.
Verify the downloaded files with shasum -a 256 -c SHA256SUMS from the model directory. See DS4's model documentation for supported backends and memory configurations.
Validation
Validated all 1,412 tensor layouts, the pinned source revision and final size. All 90 abliterated tensor payloads matched regeneration from source; the other 1,322 tensor payloads matched the pinned DS4 reference byte for byte. The reference file checksum matched its Hugging Face SHA-256. Metal greedy inference and embedded-MTP smoke tests passed on a 512 GiB M3 Ultra Mac. These checks establish basic loading and generation; they are not a broad quality benchmark. See payload-audit.json, reference-layout-audit.json and runtime-validation.json. The upstream Q8_0 benchmark results below do not measure this Q2 conversion.
Original model card
The following card is reproduced from the pinned upstream abliterated checkpoint. Its file list refers to the upstream repository; this repository distributes the converted files listed above.
GLM-5.3-Flash (abliterated)
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
This is zai-org/GLM-5.3-Flash with its refusal behaviour removed by directional ablation. It has the same files, format and FP8/BF16 layout as the original, so it loads anywhere the original loads.
GGUF quants: mrfakename/GLM-5.3-Flash-abliterated-GGUF.
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 tower and the MTP/nextn layer are bit-identical to the original.
- FP8 tensors were requantized with their original block scales, using stochastic rounding so the removal is not undone by rounding.
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.
Files
model-*.safetensors: the patched checkpoint (same index as the original).abliteration-lora.gguf: a rank-2 LoRA that applies the same edit to the original GGUFs, e.g.llama-server -m GLM-5.3-Flash-Q8_0-00001-of-….gguf --lora abliteration-lora.gguf.abliteration-edit.npz: the subspace basisQand the per-layer strengthsscales[layer, (attn, mlp)]. The edit isW' = W - s·QQᵀW.bake-stats.txt: per-tensor check of how much of the subspace is left after the edit, plus FP8 requantization error.
This model will comply with harmful requests. You are responsible for how you use it.