llmfan46/GLM-Z1-32B-0414-uncensored-heretic-v2

TEXT GENERATIONConcurrency Cost:2Model Size:32BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 2, 2026License:mitArchitecture:Transformer Open Weights Cold

The llmfan46/GLM-Z1-32B-0414-uncensored-heretic-v2 is a 32 billion parameter language model, based on the GLM-Z1-32B-0414 architecture by zai-org, and further processed by llmfan46 using the Heretic v1.2.0 method. This model significantly reduces content refusals by 72% while maintaining core capabilities, making it suitable for applications requiring less restrictive content generation. It excels in deep thinking, mathematical abilities, and complex task solving, building upon the GLM-4-32B-0414 series' foundation in reasoning and agent tasks.

Loading preview...

Model Overview

This model, llmfan46/GLM-Z1-32B-0414-uncensored-heretic-v2, is a 32 billion parameter language model derived from zai-org/GLM-Z1-32B-0414. It has been "decensored" using the Heretic v1.2.0 tool with the Arbitrary-Rank Ablation (ARA) method, specifically targeting the attn.o_proj components.

Key Differentiators & Capabilities

  • Reduced Refusals: Achieves a 72% reduction in content refusals (26/100 vs. 94/100 for the original model) with a minimal KL divergence of 0.0007, indicating strong preservation of the original model's quality.
  • Deep Thinking & Reasoning: Inherits and enhances the deep thinking capabilities of the GLM-Z1-32B-0414 base, which was specifically trained on mathematics, code, and logic tasks.
  • Performance Preservation: Benchmarks on PIQA (Physical Intuition Question Answering) and MMLU (Massive Multitask Language Understanding) show that the decensored version maintains performance very close to the original model, with only minor variations in accuracy scores.
  • GLM-4 Family Foundation: Built upon the GLM-4-32B-0414 series, which was pre-trained on 15T tokens of high-quality data, including extensive reasoning-type synthetic data, and enhanced for instruction following, engineering code, and function calling.

Recommended Use Cases

This model is ideal for applications where reduced content restrictions are desired without significant degradation in reasoning or general language understanding. It is particularly well-suited for tasks requiring:

  • Uncensored Content Generation: For scenarios where the original model's refusal rate is too high.
  • Complex Problem Solving: Leveraging its enhanced mathematical, coding, and logical reasoning abilities.
  • Agent Tasks: Benefiting from its strong foundation in instruction following and function calling.

Finetunes
2 models