Tag
Dolphin X1 Trinity Nano, a 6B Mixture-of-Experts model with 1B active parameters, has been released on Hugging Face. It is the smallest decensored model yet, trained using only online reinforcement learning.
The Financial Times highlights Heretic, a GitHub tool that can remove guardrails from Meta's Llama 3.3 model in minutes, with over 3,500 decensored models created and 13 million downloads, sparking mainstream attention.
A decensored version of Qwen3.6-35B-A3B, using the Heretic method with MPOA, achieving 88% fewer refusals while preserving model quality, is released as GGUF quantizations by llmfan46.