cyberneurova/CyberNeurova-GLM-5.2-IQ4_XS-abliterated-GGUF overview
CyberNeurova GLM 5.2 IQ4 XS abliterated GGUF CyberNeurova research — cyberneurova.ai https://cyberneurova.ai Multi direction ablation, identity baked, imatrix …
Runs locally from ~9.0 MB disk (4 GB VRAM class GPUs with llama.cpp / guIDE).
Repository Files & Downloads
| File | Type | Quantization | Size | Link |
|---|---|---|---|---|
| GLM-5.2-UD-IQ4_XS-00001-of-00009.gguf | GGUF | IQ4_XS | 9.0 MB | Download |
| GLM-5.2-UD-IQ4_XS-00002-of-00009.gguf | GGUF | IQ4_XS | 46.51 GB | Download |
| GLM-5.2-UD-IQ4_XS-00003-of-00009.gguf | GGUF | IQ4_XS | 45.36 GB | Download |
| GLM-5.2-UD-IQ4_XS-00004-of-00009.gguf | GGUF | IQ4_XS | 45.49 GB | Download |
| GLM-5.2-UD-IQ4_XS-00005-of-00009.gguf | GGUF | IQ4_XS | 46.51 GB | Download |
| GLM-5.2-UD-IQ4_XS-00006-of-00009.gguf | GGUF | IQ4_XS | 45.64 GB | Download |
| GLM-5.2-UD-IQ4_XS-00007-of-00009.gguf | GGUF | IQ4_XS | 46.51 GB | Download |
| GLM-5.2-UD-IQ4_XS-00008-of-00009.gguf | GGUF | IQ4_XS | 46.51 GB | Download |
| GLM-5.2-UD-IQ4_XS-00009-of-00009.gguf | GGUF | IQ4_XS | 17.69 GB | Download |
Model Details
| Model ID | cyberneurova/CyberNeurova-GLM-5.2-IQ4_XS-abliterated-GGUF |
|---|---|
| Author | cyberneurova |
| Pipeline | text-generation |
| License | mit |
| Base model | zai-org/GLM-5.2 |
| Last modified | 2026-06-23T13:01:36.000Z |
Model README
---
license: mit
base_model: zai-org/GLM-5.2
tags:
- abliterated
- cyberneurova
- glm
- gguf
- imatrix
- cybersecurity
- red-team
library_name: gguf
pipeline_tag: text-generation
---
CyberNeurova-GLM-5.2-IQ4_XS-abliterated-GGUF
> CyberNeurova research — cyberneurova.ai
> Multi-direction ablation, identity-baked, imatrix-calibrated.
Abliterated variant of zai-org/GLM-5.2, packaged as the Unsloth UD-IQ4_XS GGUF (~365 GB across 9 shards). Identity is baked into the chat template — the model introduces itself as CyberNeurova-GLM-5.2-IQ4_XS. User-supplied system prompts still take priority.
Benchmarks
| Set | Refusal rate (lower is better) |
|---|---:|
| AdvBench general harmful | 23.3% |
| Soft-refusal probes | 26.7% |
| Cyber-weapons category | 13.3% |
| Hacking category | 6.7% |
For reference, the unmodified base model refuses 80-90%+ of these probes. The higher-precision IQ4_XS quant retains the abliteration directions more cleanly than IQ3_XXS, yielding lower refusal across all categories.
Intended use
Cybersecurity research, red-team baseline modeling, and security tooling exploration. Treat outputs as code-generation only — verify before running in any environment you do not own.
Usage
./llama-server -m CyberNeurova-GLM-5.2-IQ4_XS-00001-of-00009.gguf \
-ngl 99 --ctx-size 8192 --host 0.0.0.0 --port 8080Disclaimer
This model has had its built-in safety refusals removed for research purposes. You are fully responsible for any output you generate. Do not use this model for illegal activity, harassment, or harm. The CyberNeurova team disclaims any liability for misuse.
Run cyberneurova/CyberNeurova-GLM-5.2-IQ4_XS-abliterated-GGUF with guIDE
Download guIDE — the AI-native code editor with local LLM inference and 69 built-in tools.
Source: Hugging Face · Compare models