arnomatic/MetalGPT-1-heretic

Hugging Face
TEXT GENERATIONConcurrency Cost:2Model Size:32BQuant:FP8Ctx Length:32kPublished:Dec 15, 2025License:apache-2.0Architecture:Transformer Open Weights Warm

arnomatic/MetalGPT-1-heretic is a 32 billion parameter decensored version of the nn-tech/MetalGPT-1 model, built upon the Qwen/Qwen3-32B architecture. This model has been 'abliterated' using Heretic v1.1.0 to significantly reduce refusals, making it suitable for use cases requiring less restrictive content generation. It retains the original model's domain-specific fine-tuning for the mining and metallurgy industry, excelling in Russian-language technical discussions within this field.

Loading preview...

What the fuck is this model about?

arnomatic/MetalGPT-1-heretic is a 32 billion parameter language model derived from the nn-tech/MetalGPT-1 base, which itself is built on the Qwen/Qwen3-32B architecture. Its core purpose is to provide a decensored version of the original model, achieved through a process called "abliteration" using the Heretic v1.1.0 tool. This modification drastically reduces content refusals, making it more permissive in its responses.

What makes THIS different from all the other models?

The primary differentiator is its decensored nature. While the original MetalGPT-1 model had a refusal rate of 100/100 on tested prompts, this 'heretic' version achieves a significantly lower refusal rate of 14/100. This makes it uniquely suited for applications where the base model's safety filters might be overly restrictive. Furthermore, it retains the original model's specialized training in the mining and metallurgy industry, making it proficient in Russian-language technical discussions within this domain.

Should I use this for my use case?

  • Use this model if:

    • Your application requires a model with significantly reduced content refusal rates.
    • You need a powerful 32B parameter model for technical discussions in Russian, specifically within the mining and metallurgy sectors.
    • You are working on tasks that might trigger safety filters in more heavily moderated models, but where the content is contextually appropriate for your domain.
  • Consider alternatives if:

    • You require a general-purpose language model without specific domain expertise.
    • Your application strictly requires adherence to strong safety guidelines and content moderation.
    • You are not working with Russian language or the mining/metallurgy domain.