What are GGUF, MLX, fine-tune and abliterated versions?
Who made each copy of a model, what changed in it and which one to download.
Short answer
GGUF and MLX are file formats: the same model, packaged for different apps. Fine-tune and abliterated are changed models: someone altered the original.
- GGUF: One file that Ollama, LM Studio and llama.cpp open on any computer.
- MLX: Apple's format for Macs with an M chip, opened by LM Studio.
- Fine-tune: The original trained further on new data. A different model.
- Abliterated: A copy edited by a third party so it refuses less.
Formats: (Hugging Face · GGUF, read Oct 17, 2026) · (Apple · MLX, read Oct 17, 2026)
Kinds of versions
- Original
- The weights as the publisher trained themUsually 16-bit (BF16) safetensors files, meant for servers and for further training.
- Official quantized
- A smaller copy released by the publisherFP8, NVFP4 or AWQ files, mostly for server GPUs.
- GGUF
- One file per precision for llama.cppConverted by the publisher or by the community.Hugging Face · GGUF
- MLX
- A copy for Apple's MLX frameworkConverted by the community, mostly for Macs.Apple · MLX
- Fine-tune
- The model trained further on new dataChanges what it knows or how it answers. Same size and memory as the original.
- Adapter · merge
- A small add-on file, or two models blendedAn adapter (LoRA) loads on top of the original; a merge combines the weights of several models.
Which of the two to download on a Mac or in LM Studio, with the apps that open each one, real sizes and measured speeds: GGUF vs MLX.
One model, four copies
Qwen3.6 27B on Hugging Face. Sizes summed from each repository's weight files, read Oct 16, 2026.
| Version | Repository | Made by | Size (GB) |
|---|---|---|---|
| Original · BF16 | Qwen/Qwen3.6-27B | Publisher | 55.6 |
| Official quantized · FP8 | Qwen/Qwen3.6-27B-FP8 | Publisher | 30.9 |
| GGUF · Q4_K_M | unsloth/Qwen3.6-27B-GGUF | Community | 16.8 |
| MLX · 4-bit | lmstudio-community/Qwen3.6-27B-MLX-4bit | Community | 16.1 |
The four copies share one set of parameters, so the site plans them with one estimate per precision (bytes per parameter). Benchmark scores on this site are for the publisher's original weights; a converted copy can score a little differently, and a fine-tune is a different model.
How to tell them apart
On Hugging Face, a copy names its original in the base_model field, and the Hub labels the link as a fine-tune, an adapter, a quantized version or a merge. The copy's page then shows its base model (Hugging Face · base_model, read Oct 17, 2026).
- Who uploaded it
- The publisher's account, or someone else'sQwen/… is the publisher; unsloth/… and lmstudio-community/… are converters.
- What the name adds
- GGUF, MLX, a bit count, or a new nameA new name usually means a fine-tune or a merge: a different model.
- Licence
- Read the original's licence tooCopies are built from the original weights and may be bound by its terms.
Abliterated and uncensored
Chat models are trained to decline some requests. Researchers found that this refusal behaviour is tied to one direction in the model's internal activations (Arditi et al., 2024, read Oct 17, 2026). “Abliteration” removes that direction from the weights, so the copy refuses far less often. “Uncensored” is a broader label for copies changed to refuse less, by abliteration or by fine-tuning.
These copies are made by third parties, not by the publisher. The publisher's safety testing and published results describe the original, not the copy, and its answers and quality can differ. This site lists original models only and does not carry their scores over to modified copies.