Mantis2024 drops Dirty Shirley uncensored LLM in 4-bit quantization

Dirty Shirley LoRablated Erotica Advanced v1 NSFW, a new uncensored language model from mantis2024, arrived on HuggingFace on June 18 in IQ4_XS quantized format. The model card lists transformers, GGUF, and llama-cpp tags, signaling compatibility with the local inference toolchain practitioners use to run unrestricted models on consumer hardware. The "lorablated" naming pattern indicates a merge or fine-tune that removes safety alignment layers—a technique now standard for NSFW-capable releases in the open-weight LLM space.

The IQ4_XS quantization scheme compresses weights to roughly 4 bits per parameter, a middle ground between 8-bit formats that preserve fidelity and 3-bit extremes that trade quality for speed. Quantized GGUF files are designed to run in llama.cpp, the C++ inference engine powering Ollama, LM Studio, and dozens of local LLM frontends. The model card omits parameter count, context length, and base architecture—details practitioners need for memory planning and performance tuning. Without benchmark numbers, sample outputs, or a linked base model card, it is difficult to assess whether the abliteration preserved coherence or introduced the repetition and drift that often plague safety-removal fine-tunes.

The mantis2024 namespace has not appeared in prior high-profile releases, suggesting this may be an early experiment or a niche merge targeting a specific use case. Practitioners watching the space should look for follow-up releases that include context length, parameter count, and at least one sample conversation—details that would help users decide whether the model fits their workflow. A full-precision or FP16 upload would also serve those who want unquantized weights for further merging or LoRA training.