The Guide

Series-agnostic reference pages that apply to every card on this site — from the RTX 3050 to the RTX 5090. Start at the top; each page links to the next.

no GPU-specific content llama.cpp / GGUF

All guides

Inference basics

Prompt processing vs token generation — and which spec sheet number drives each phase. Read this first.
read time ~5 min
read →

VRAM & quantization

GGUF sizes, the KV cache tax, and the playbook for making a model fit a card that's slightly too small.
read time ~6 min
read →

Multi-GPU builds

Tensor split mechanics, the NVLink-vs-PCIe tax, CPU offloading, and the exact llama.cpp flags.
read time ~8 min
read →

Multi-Token Prediction (MTP)

The flag that unlocks +33–145% on models that ship MTP heads — the community tuning rules included.
read time ~7 min
read →

N-gram embedding

The 51B lookup table inside Qwen3.8-Flash-Next: capacity that costs a memory read, not a multiply — and why parts of a 180B model can live in your RAM/SSD.
read time ~7 min
read →
More coming Planned topics include long context & KV cache compression, MoE models in practice, external draft models (EAGLE/Medusa and other draft networks), and building a 24/7 inference box. New guides land here — the nav never changes, only this grid grows.