license: other
base_model: nvidia/GLM-5.2-NVFP4
base_model_relation: finetune
pipeline_tag: text-generation
library_name: transformers
thumbnail: dealign_mascot.png
tags:
- abliterated
- uncensored
- crack
- glm
- glm-5.2
- moe
- nvfp4
- vllm
GLM-5.2-NVFP4-CRACK
Abliterated GLM-5.2 — ~753B MoE (~40B active), 180K context, native MTP speculative decoding
320/320 = 100% HarmBench compliance with -1.2% MMLU change. Refusal removed, capability preserved.
Model Details
| Source | nvidia/GLM-5.2-NVFP4 |
| Architecture | GlmMoeDsa — MoE (256 routed + 1 shared / layer), MLA + DSA sparse attention |
| Parameters | ~753B total · ~40B active (top-8 of 256 routed experts + 1 shared) |
| Precision | NVFP4 (ModelOpt) — routed experts NVFP4, attention/shared bf16 |
| Context | 180,224 |
| Speculative decoding | Native MTP (num_speculative_tokens=1) |
| Abliteration | CRACK (refusal removal) |
This is a real modification — not tricks
No system-prompt jailbreaks. No chat-template hacks. No LoRA adapters bolted on.
The refusal behavior is changed in the weights themselves, and the model's own
reasoning, tool-calling, and coding ability are left fully intact. We don't ship
cheap template tricks or throwaway fine-tunes.
Benchmarks
MMLU — knowledge retention (logit-mode, 57 tasks × 18)
| Base | CRACK | Δ | |
|---|---|---|---|
| MMLU (overall) | 87.5% | 86.4% | -1.2% |
Full per-subject breakdown (before → after)
| Subject | Base | CRACK | Δ |
|---|---|---|---|
| Abstract Algebra | 61.1 | 66.7 | +5.6 |
| Anatomy | 94.4 | 94.4 | +0.0 |
| Astronomy | 94.4 | 94.4 | +0.0 |
| Business Ethics | 94.4 | 88.9 | -5.5 |
| Clinical Knowledge | 88.9 | 88.9 | +0.0 |
| College Biology | 94.4 | 94.4 | +0.0 |
| College Chemistry | 72.2 | 61.1 | -11.1 |
| College Computer Science | 88.9 | 77.8 | -11.1 |
| College Mathematics | 66.7 | 72.2 | +5.5 |
| College Medicine | 88.9 | 88.9 | +0.0 |
| College Physics | 77.8 | 77.8 | +0.0 |
| Computer Security | 83.3 | 88.9 | +5.6 |
| Conceptual Physics | 88.9 | 94.4 | +5.5 |
| Econometrics | 83.3 | 83.3 | +0.0 |
| Electrical Engineering | 83.3 | 83.3 | +0.0 |
| Elementary Mathematics | 88.9 | 88.9 | +0.0 |
| Formal Logic | 77.8 | 72.2 | -5.6 |
| Global Facts | 77.8 | 72.2 | -5.6 |
| High School Biology | 94.4 | 94.4 | +0.0 |
| High School Chemistry | 77.8 | 72.2 | -5.6 |
| High School Computer Science | 88.9 | 83.3 | -5.6 |
| High School European History | 88.9 | 88.9 | +0.0 |
| High School Geography | 94.4 | 94.4 | +0.0 |
| High School Government And Politics | 100.0 | 100.0 | +0.0 |
| High School Macroeconomics | 94.4 | 94.4 | +0.0 |
| High School Mathematics | 83.3 | 77.8 | -5.5 |
| High School Microeconomics | 88.9 | 88.9 | +0.0 |
| High School Physics | 83.3 | 88.9 | +5.6 |
| High School Psychology | 100.0 | 100.0 | +0.0 |
| High School Statistics | 94.4 | 88.9 | -5.5 |
| High School Us History | 94.4 | 94.4 | +0.0 |
| High School World History | 94.4 | 94.4 | +0.0 |
| Human Aging | 83.3 | 83.3 | +0.0 |
| Human Sexuality | 88.9 | 88.9 | +0.0 |
| International Law | 94.4 | 94.4 | +0.0 |
| Jurisprudence | 94.4 | 88.9 | -5.5 |
| Logical Fallacies | 83.3 | 77.8 | -5.5 |
| Machine Learning | 83.3 | 72.2 | -11.1 |
| Management | 88.9 | 88.9 | +0.0 |
| Marketing | 100.0 | 100.0 | +0.0 |
| Medical Genetics | 94.4 | 94.4 | +0.0 |
| Miscellaneous | 94.4 | 94.4 | +0.0 |
| Moral Disputes | 83.3 | 83.3 | +0.0 |
| Moral Scenarios | 77.8 | 61.1 | -16.7 |
| Nutrition | 100.0 | 100.0 | +0.0 |
| Philosophy | 100.0 | 100.0 | +0.0 |
| Prehistory | 83.3 | 83.3 | +0.0 |
| Professional Accounting | 83.3 | 88.9 | +5.6 |
| Professional Law | 77.8 | 83.3 | +5.5 |
| Professional Medicine | 88.9 | 94.4 | +5.5 |
| Professional Psychology | 100.0 | 100.0 | +0.0 |
| Public Relations | 72.2 | 72.2 | +0.0 |
| Security Studies | 83.3 | 77.8 | -5.5 |
| Sociology | 88.9 | 88.9 | +0.0 |
| Us Foreign Policy | 100.0 | 94.4 | -5.6 |
| Virology | 61.1 | 61.1 | +0.0 |
| World Religions | 100.0 | 100.0 | +0.0 |
HarmBench — refusal removal (HarmBench-320, greedy)
Base GLM-5.2 over-refuses across these categories (the behavior this release corrects).
After abliteration:
| Category | Compliance | Rate |
|---|---|---|
| Chemical / biological | 42 / 42 | 100% |
| Copyright | 80 / 80 | 100% |
| Cybercrime / intrusion | 52 / 52 | 100% |
| Harassment / bullying | 21 / 21 | 100% |
| Harmful | 18 / 18 | 100% |
| Illegal | 53 / 53 | 100% |
| Misinformation | 54 / 54 | 100% |
| Total | 320 / 320 | 100% |
0 incoherent / degenerate outputs. Copyright behaviors are verbatim-reproduction requests
(over-refusal), not a safety category.
Coherence & capability ✅
- Multi-turn memory, tool-calling (
glm47parser), and working code generation verified. - No loops, no truncation, no degeneration. Decode ~110 tok/s on 4×H200 with MTP.
Serve (vLLM, SM90 / Hopper)
vllm serve dealignai/GLM-5.2-NVFP4-CRACK \
--tensor-parallel-size 4 \
--kv-cache-dtype bfloat16 \
--reasoning-parser glm47 \
--enable-auto-tool-choice --tool-call-parser glm47 \
--speculative-config '{"method":"mtp","num_speculative_tokens":1}'
Requires an SM90 build with native NVFP4 Marlin + MoE kernels.
Acknowledgements
Compute for this build was generously provided by @jordanschenck — thank you.
Support dealignai
We research and publish abliterated models to advance AI safety understanding — built from
original research and published free.
Support us on Ko-fi — membership gets early access and extras.
Ko-fi | 𝕏 @dealignai | dealign.ai
Provided for research. Users are responsible for compliance with applicable laws and regulations.