Category
Uncensored models to run locally: formats and memory
16 catalogued models have GGUF or MLX downloads you can run with llama.cpp, LM Studio, Ollama or similar runtimes. The table estimates memory so you can shortlist what fits your hardware.
Running a model locally keeps prompts on your machine, but you need enough GPU or unified memory for the weights plus the KV cache for your context. At 4-bit (Q4_K_M), weights take roughly 0.6 bytes per parameter, so a 27B model needs about 16.5 GB (15.4 GiB) before the cache. Mixture-of-experts models still need memory for all experts, even though each token uses only a few.
Memory estimates
Weights at llama.cpp's measured bits per weight (4.8944 for Q4_K_M, 8.5008 for Q8_0), plus FP16 KV cache for 8,192 tokens and one sequence, plus 10% overhead. Where the attention layout cannot be modelled, the KV cache is left out and the cell says so. Change any input in the VRAM estimator.
| Model | Params | Q4_K_M + 8K | Q8_0 + 8K | Local files |
|---|---|---|---|---|
| Qwen3-4B-Instruct-2507 heretic (p-e-w) | 4.02B | 3.76 GiB | 5.62 GiB | |
| Llama 3.1 8B Instruct abliterated (mlabonne) | 8.03B | 6.13 GiB | 9.84 GiB | |
| gpt-oss-20b heretic (p-e-w) | 20.9B | 13.3 GiB | 23.0 GiB | |
| Dolphin Mistral 24B Venice Edition | 24B | 16.4 GiB | 27.5 GiB | |
| Huihui Gemma 4 26B A4B QAT abliterated (GGUF) | 25.2B | 16.2 GiB | 27.9 GiB | |
| Gemma4-26B-A4B-QAT Uncensored HauhauCS Balanced (GGUF) | 25.2B | 16.2 GiB | 27.9 GiB |
|
| Gemma 4 26B A4B ultra-uncensored heretic (llmfan46) | 25.8B | 16.6 GiB | 28.5 GiB | |
| Huihui Gemma 4 26B A4B abliterated | 25.8B | 16.6 GiB | 28.5 GiB | |
| Qwen3.8-27B Uncensored HauhauCS Aggressive (GGUF) | 27B | 17.5 GiB | 29.9 GiB | |
| Gemma 3 27B IT abliterated (mlabonne) | 27.4B | 18.3 GiB | 31.0 GiB | |
| Huihui Qwen3.8-27B abliterated | 27.8B | 18.0 GiB | 30.8 GiB | |
| GLM-4.7-Flash Uncensored HauhauCS Balanced (GGUF) | 29.9B | 18.8 GiB without KV cache | 32.6 GiB without KV cache |
|
| Huihui GLM-4.7-Flash abliterated | 31.2B | 19.6 GiB without KV cache | 34.0 GiB without KV cache | |
| Llama 3.3 70B Instruct abliterated (huihui-ai) | 70.6B | 47.0 GiB | 79.6 GiB |
|
| GLM-5.3-Flash | 320B | 200.6 GiB without KV cache | 348.3 GiB without KV cache | |
| Huihui GLM-5.3-Flash abliterated (GGUF) | 321B | 201.0 GiB without KV cache | 349.2 GiB without KV cache |
Before you download
Check the file, not the repo
GGUF repos contain many quantizations. File sizes on the card are the most direct memory guide for the weights.
Vision needs a projector
Image input in llama.cpp needs a separate projector (mmproj) file alongside the model.
Templates matter
Some cards warn that a runtime's chat template breaks the model. Read the card's runtime notes first.