Granite 3.3 2B abliterated weights land on HuggingFace in GGUF format
An uncensored version of IBM's Granite 3.3 2B instruction model appeared on HuggingFace this week, removing safety layers from the small conversational checkpoint.
Granite 3.3 2B Uncensored, released by community creator arzaan789, is an abliterated text-generation model that strips safety guardrails from IBM's Granite 3.3 2B Instruct checkpoint. The weights landed on HuggingFace on May 13 in GGUF format, targeting local deployment for practitioners who need unrestricted conversational output.
The base Granite 3.3 2B Instruct is IBM's small instruction-tuned variant, designed for dialogue and task completion on consumer hardware. Abliteration removes the safety alignment layers added during instruction tuning, leaving the underlying language model intact but eliminating refusal behavior. At 2 billion parameters, it sits in the same weight class as Phi-3 Mini and Gemma 2 2B—small enough to run on a laptop GPU but large enough for coherent multi-turn conversation.
Deployment and licensing
The GGUF packaging enables quantized inference via llama.cpp or similar runtimes, making it portable across consumer devices. The model inherits IBM's Apache 2.0 license from the base Granite checkpoint, permitting commercial use without restriction. Practitioners running local stacks can pull the weights directly from the repository. The small parameter count and GGUF format make it a candidate for edge devices, though real-world performance will depend on the quantization level and whether the abliteration process preserved coherence across the full capability range of the base model.