File size: 1,621 Bytes
42b2ae6 f2e20a8 42b2ae6 f2e20a8 bbf74ed f24cfd6 f2e20a8 f24cfd6 f2e20a8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 | ---
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.*
|