Getting Started
Where to start.
A suggested walk through the site for someone who has never run a local model before, one lever at a time, in the order that actually holds up rather than the order things were written in.
Maintained rather than published: stations move from "coming soon" to linked the moment the article behind them ships, and the date above moves with it. If a station feels out of order, say so.
Hardware and the case for it first, then the three levers that govern what will actually run (architecture, quantisation, context), then the tool that turns those levers into a shortlist, then the stack that puts a model in front of you day to day.
Your path through Dev Factory
9 stops · 6 live now · scroll to move the marker
The case for going local at all: what renting model access actually exposes you to, and why consumer hardware is finally good enough.
What to actually buy, tier by tier, now that both vendors have quietly dropped their budget cards from active toolkit support.
Getting the box itself working, drivers and containers included, before anything runs on it.
The first lever: total parameters versus active parameters, and why a mixture-of-experts model can run faster than its size suggests.
The second lever: how fewer bits shrink model weights, and why a label such as Q4_K_M cannot guarantee quality or speed on its own.
The third lever: why the context window on a model card and the one you actually get are not the same claim.
Where the first three levers meet: turning total parameters and a quantisation choice into a shortlist for the GPU you already own, planner built in.
The local-ChatGPT path: the first thing you will actually run day to day, and when it stops being enough.
From "a model is running" to "a model is working for me": the bridge into agentic tooling.
Going further
This path stops at a working, agent-capable setup on purpose. Once it feels solid, the deeper architecture pieces reward knowing why quantisation and VRAM budgets matter first: KV cache and the hybrid-architecture ROCm finding, the engine landscape across five different runtimes, and the multi-GPU agentic platform the fleet actually runs. None of them are prerequisites for a working setup, they are what comes after one.
Related
- Hardware: what is actually racked and running on this fleet, referenced throughout the path above.
- Configs: the downloadable launch scripts and compose files the later stations point back to.
- Models: the fleet's own picks by VRAM tier, for choosing what to actually run once this path is done.
- Resources: the living list of projects, tools and channels this path draws on.