🤖 New Models for Coding and Agentic Behavior on Hugging Face
An 'abliterated' version of Gemma-4-12B-coder by huihui-ai has been released, in which safety refusal mechanisms were removed using the abliteration method. Additionally, a LoRA adapter for Qwen3.6-27B has been presented, trained on the Glint-Research/Fable-5-traces dataset, transforming the model into an autonomous agent.
🌍 The development of 'abliteration' methods and CoT (Chain of Thought) distillation from closed models allows for the creation of powerful open-weights models with agentic capabilities, bypassing standard safety constraints.
👤 It is now possible to use more efficient models for development automation that better understand complex instructions and are less likely to refuse tasks.
Source 1: https://huggingface.co/huihui-ai/Huihui-gemma-4-12B-coder-fable5-composer2.5-v1-abliterated Source 2: https://huggingface.co/hotdogs/qwen3.6-27b-fable5-lora