Real benchmarks. Real hardware. Real costs.
Run AI at home, practically.
GPU benchmarks, local LLM tutorials, ComfyUI workflows, and self-hosted AI guides for the home lab crowd. Tested on the cards you can actually buy.
What we cover
Four corners of the home AI lab.
-
GPU Benchmarks
Real numbers for Stable Diffusion, Flux, ComfyUI, and local LLMs across the cards normal humans can afford.
-
Local LLM Tutorials
Step-by-step Ollama, llama.cpp, LM Studio, and quantized model picks that fit your VRAM, not a datacenter's.
-
ComfyUI & Image Gen
Workflows that work, LoRA training on limited VRAM, and node setups without the YouTube guesswork.
-
Self-Hosted Stack
Open WebUI, n8n automations, privacy-first infra. Your data, your hardware, your rules.
Latest posts
Fresh from the bench.
-
Liquid AI LFM3 for Local AI: Six 2B Models, a 1.1GB GGUF, and Why You Don't Need New Hardware to Run Them (2026)
LFM3 puts text, vision, audio, and decision models in ~2B parameters each. Real tok/s numbers, the $10M license catch, and the Ollama error to expect.
Read guide → -
Survival LLM Hardware Guide 2026: The Offline Emergency AI Is a RAM Ladder — 55% Accuracy at 16GB, 77% at 32GB, and No Raspberry Pi Option
Survival LLM runs an offline emergency library plus a local model sized to your RAM. Its own benchmark shows why 32GB is the real floor — and why a Pi can't run it.
Read guide → -
MiniMax M3.1 Flash Open Weights: The 428B Model That Ships Pre-Quantized, and the 150GB Line That Decides Who Runs It (2026)
MiniMax M3.1 Flash weights are live: 428B/23B-active MoE, NVFP4 experts out of the box, 143.9GB at 2-bit. What actually fits a Mac Studio or home rig.
Read guide → -
Voxtral 3 for Local AI: Mistral Shipped No Weights — Here's What You Can Actually Run (2026)
Voxtral 3 is API-only. The open-weight Voxtral models you can still run locally: VRAM requirements, GGUF sizes, which GPU fits, and when the API wins.
Read guide → -
llama.cpp v0.7.0 MoE Cache: What the 2.2x Decode Claim Actually Delivers on Your GPU (2026)
llama.cpp v0.7.0 hands-on: real --moe-cache-mib benchmarks on RTX 4090/5090, who gains nothing, the 8080-to-9931 port change, and the LiliCorr reality.
Read guide →