Read this first
These models have no guardrails, so you own every answer they give. The refusal numbers below are each releaser's own test, not independent results.
The sizes are file sizes. You need that much free memory plus extra room for context. Some repos are gated: open the page and accept the terms once, then download.
Install in 3 steps
-
Install LM Studio or Ollama
Either one runs these models on your own computer. LM Studio is the easier one if you like clicking over typing.
-
Pick your memory tier below
Match the tier to the memory on your PC or Mac. On a Mac, that is unified memory.
-
Download the model
Tap a model, copy its name from Hugging Face, and search for it inside LM Studio or Ollama. Start with the first one in your tier.
32 GB
One RTX 5090, Radeon AI Pro R9700, or two 16 GB cards split.
- Qwen 3.6 40B Deckard Heretic40B thinking merge, Q5, 28.2 GB
- Qwen 3.8 27B AggressiveQ6, 25.9 GB, 0 refusals out of 465
- Gemma 4 31B HereticQ6, 25.2 GB, refusals cut from 99 to 10 out of 100
- Qwen 3.6 35B A3B AggressiveQ6, 30.6 GB, 0 refusals out of 465
Three 3060s (36 GB) fit Gemma 4 31B Heretic at Q8 (32.6 GB).
48 GB
2x RTX 3090, 2x RTX 4090, RTX A6000 48 GB, RTX Pro 5000, or 2x 7900 XTX.
96 to 128 GB
RTX Pro 6000 Blackwell 96 GB, AMD Strix Halo 128 GB, DGX Spark 128 GB, or Mac Studio 128 GB.
256 GB
2x DGX Spark or Mac Studio 256 GB.
512 GB
4x DGX Spark or Mac Studio 512 GB. The deep end.
Where this list comes from
This list was put together by sudoingX on X. All credit to him for mapping the models.
See the original postAMD cards run the GGUF files on llama.cpp (Vulkan or ROCm). Macs run them on llama.cpp, or the MLX builds where they exist.