Qwen3-24B abliterated weights hit HuggingFace with 256k context
Higashikawa released abliterated GGUF weights for Qwen3-24B-A4B-Freedom-Thinking, a mixture-of-experts model with 256k context and no safety filters.
Higashikawa released abliterated GGUF weights for Qwen3-24B-A4B-Freedom-Thinking on HuggingFace. The model is a mixture-of-experts architecture built on Qwen3, with 24 billion parameters split across six 4-billion-parameter experts and a 256k token context window. The "Abliterated-Heretic-NEO" designation signals removal of safety guardrails, making it an uncensored variant of the base Qwen3 model.
The GGUF format means the weights are quantized for local inference on consumer hardware. The model card lists "all use cases" as supported, with no content restrictions. The mixture-of-experts design activates a subset of parameters per token, which typically reduces compute cost compared to dense models of similar total size while maintaining performance on reasoning tasks.
What stands out
- 01256k context window — among the longest in open-weight models at this parameter count, enabling document-scale reasoning without chunking.
- 02Mixture-of-experts at 24B — six 4B experts means roughly 8B active parameters per forward pass, cutting inference cost while preserving the model's knowledge footprint.
- 03Abliterated weights — safety filters removed, making it a true uncensored alternative to the base Qwen3 release.
- 04GGUF quantization — ships ready for llama.cpp, Ollama, and other local inference tools without additional conversion.
