metadata
title: README
emoji: 🔥
colorFrom: yellow
colorTo: red
sdk: static
pinned: false
Hob Forge
Small household spirits for small machines.
A hob is an English household spirit that quietly does the work. Hob Forge exists for one reason: to help you actually run good models on the hardware you already have. We build 3B–4B finetunes and merges — the best little agentic models — and we are unapologetically the champion of small GPUs and modest RAM: every release states exactly what it needs at 8GB, 12GB, and 16GB (and what CPU-only costs you), and ships with the exact way to run it — step-by-step guides, tested configs, real support.
Every release promises:
- A stated memory budget. Quant sizes and KV-cache maths per hardware tier — 8GB, 12GB, 16GB — in every card, measured, not projected. If it doesn't fit with real context at a tier, the card says so plainly.
- Honest evals. Held-out benchmarks only, decontamination stated, baselines included — the raw base and a same-size peer, same harness, same settings. We publish the harness output, not vibes.
- Full provenance. Which teachers, which datasets, which licenses, what filtering — disclosure tables with verified-trace counts.
- A job, not a persona. Each hob does one thing well — terminal work, code fixing, tool routing. Merges combine proven specialists, and a merge that doesn't beat its parents doesn't get a name.
First resident in the works: smeagle — a 4B terminal/software-engineering agent.
Built in a home lab in the UK on a strict diet of verified data.