Specs
Triple boot- GPU
- AMD Radeon RX 6800 (16 GB)
- CPU
- AMD Ryzen 9 3900X (12 cores / 24 threads)
- RAM
- 32 GB
- Triple boot
About
I’m an AI enthusiast and I self-host everything. I run large models locally on my own hardware, and I’m particularly interested in abliterated models.
My main PC triple-boots Windows 11, Arch Linux and a Hackintosh. My home server dual-boots Proxmox and another Hackintosh. So that’s two PCs that spend part of their lives pretending to be Macs.
I like owning the whole stack: the hardware, the operating systems on it, and the models running on top.
Main rig
One PC, three operating systems. One of them thinks it’s a Mac.
Home server
The second machine, in the same format. It dual-boots, and one side of it is a Mac that isn’t.
Specs
Dual boot- GPU
- Integrated graphics (Intel HD Graphics 4600)
- CPU
- Intel Core i5-4570 (4 cores / 4 threads)
- RAM
- 8 GB
- Storage
- 256 GB
- Dual boot
- Runs
Local AI
Everything here runs on the main rig, on my own hardware.
- Runtime
- llama.cpp on the Vulkan backend, about 30 tokens a second on the RX 6800.
- Front end
- Open WebUI, plus Open Terminal.
What 30 tokens a second looks like
This is roughly how fast a 27B model writes on my RX 6800, running llama.cpp on the Vulkan backend: about thirty tokens every second, which is faster than most people read.
Abliterated models
Chat models are trained to refuse some requests, and it turns out that refusal mostly comes down to one direction inside the model. Abliteration finds that “refusal direction” and removes it from the weights, so the model stops refusing without being retrained.
Models
Qwen3.8-27B GSQ-RCO GGUF
by ISTA-DASLab
Non-uniform, mixed-precision quantization. IQ3_S is their “task-lossless” point at around 12 GB, and I run it with the MTP head (multi-token prediction) for speculative decoding.
View on Hugging FaceHuihui Qwen3.8-27B abliterated GGUF
by huihui-ai
A refusal-removed variant of the same model, made with abliteration.
View on Hugging Face