Abliteration / refusal ablation
A technique that analyzes and modifies refusal-related directions in model weights to reduce refusals. Results vary by model and implementation, with possible residual refusals and capability loss. See Refusal in Language Models Is Mediated by a Single Direction.
Uncensored
A broad community label that may involve fine-tuning, ablation, data selection or a combination of methods. It is not a standardized evaluation.
Heretic
An automated model ablation tool. We identify this category from publisher names and tags without independently verifying the configuration. See the Heretic project.
GGUF
A model file format used in the llama.cpp ecosystem, supporting various quantization levels for local CPU, GPU or hybrid inference. The format does not indicate capability.
Quantization
Reducing weight precision to save memory and computation. Labels such as Q4 and Q8 roughly describe bit width; file size also depends on architecture and quantization scheme.
MLX
A machine learning framework and model ecosystem for Apple silicon. Apple unified memory can support inference but is not directly equivalent to discrete GPU VRAM.
Base model
The upstream source of a model: a pretrained, instruction-tuned or intermediate fine-tuned model. A quantized repository may name another quantized or ablated repository as its base.
B and MoE
B means billion parameters. MoE stands for mixture of experts. For example, 35B-A3B usually means about 35B total parameters with 3B active per token. Memory for all weights still depends mainly on the total parameter count.
VRAM and KV cache
Inference memory includes weights, runtime buffers and KV cache. Long contexts and concurrency grow the KV cache, so parameter-based estimates cannot guarantee a model will fit.
Downloads
We sum the last-30-day repository download statistics returned by Hugging Face. Downloads are neither a performance ranking nor a unique-user count.