The decisions, explained
Each guide answers one question that people actually ask before they build a local model — and states its answer in the first paragraph rather than the last.
RAG vs fine-tuning
Should I use retrieval or train the model?
When retrieval beats training, when training beats retrieval, and the specific numbers that decide it. The short version: most people who think they need a fine-tune need RAG.
Read →LoRA vs QLoRA
Which fine-tuning method should I use?
LoRA freezes the base model in half precision; QLoRA freezes it in 4-bit. What that costs, what it saves, and which one your GPU can actually run.
Read →Quantization explained
Q4 or Q8 — what does quantization actually cost?
What the GGUF quantization names mean, the real bits per weight behind them, and how to choose one for your hardware without destroying the model.
Read →How much VRAM do you need?
What can I run on my GPU?
The arithmetic behind local model memory: weights, KV cache and runtime overhead, and what each common VRAM budget actually buys you.
Read →What makes a good training dataset
Why did my fine-tune not work?
The dataset problems that quietly ruin a fine-tune — duplicates, inconsistent shape, empty responses, leaked personal data — and how ModelLM detects each one.
Read →Answers generated from the catalogue
These pages are built from the same model and hardware data the product uses, so they stay correct as the catalogue changes rather than going stale the day they were written.