Wizard-Vicuna-7B-Uncensored GGUF quantized weights land on HuggingFace
B1ackJesus published GGUF-quantized weights for Wizard-Vicuna-7B-Uncensored, a 7-billion-parameter Llama-based model trained on 70,000 unfiltered instruction pairs.
B1ackJesus published GGUF-quantized weights for Wizard-Vicuna-7B-Uncensored on HuggingFace. The model is a 7-billion-parameter Llama derivative trained on the ehartford/wizard_vicuna_70k_unfiltered dataset — 70,000 instruction-response pairs with no content filtering. The GGUF format targets CPU and low-VRAM inference via llama.cpp and compatible runtimes, making the weights accessible to practitioners running local inference on consumer hardware; quantized variants typically fit in 4–8 GB of RAM depending on bit depth.
The base model combines Wizard-LM instruction tuning with Vicuna's conversational fine-tuning. The unfiltered training set includes prompts that would typically trigger safety refusals in commercial models. The model card lists an "other" license, which means users need to check the underlying quixiai base model's terms before deploying in production.
The exact GGUF quantization levels included — Q4_K_M, Q5_K_S, Q8_0, or a full suite — remain to be confirmed by checking the repository directly. File sizes can be cross-referenced against llama.cpp's quantization table to determine which variants are available. If the conversion is clean, Wizard-Vicuna-7B-Uncensored should integrate into any workflow already handling 7B Llama GGUF files, though practitioners should verify the license terms with the base model maintainers before production use.
