Starter Kit · One-Time Purchase
The Local AI Build Kit
Buy the right hardware, then get every model running on it
Worth $99 — you pay $49 today.
Available soonInstant delivery — download from your in-app library right after checkout. Linked to your account; sign in any time to re-download. All sales final.
Need help getting it set up? Contact support — we're always here to help.
Overview
Most people lose money on their first local-AI machine in one of two ways: they buy a card that can't hold the model they actually wanted, or they buy far more card than the model needs and gain nothing. Both mistakes come from the same confusion — VRAM decides WHAT runs, bandwidth and compute decide HOW FAST. This kit separates the two and answers, for your specific budget and card, what to buy and what will run on it. Every footprint and speed in it was measured on hardware, not calculated from parameter count.
What's included
- Five complete builds, from $899 CPU-only to a $3,499 RTX 4090 machine — full parts lists with measured tokens/sec on Llama 3.1 8B, Llama 3.1 70B and Mixtral 8x7B, not vendor figures.
- 72-row model-fit matrix: every model we have verified, grouped by the card it runs on, with real footprint, real speed, and a named winner per tier.
- 31 documented won't-fit traps — the model/card combinations that look reasonable and will never load, so you don't spend an evening on a 20 GB download.
- 51 tuning notes tied to specific VRAM tiers: context length, quantisation choice, KV-cache type, layer offload.
- Drop-in Ollama configs for 6, 8, 12, 16, 24 and 32 GB cards plus Apple Silicon, with two annotated Modelfiles.
- Copy-paste install commands per card, including the parts people get wrong (CUDA, ROCm, unified-memory limits on macOS).
- An OOM playbook that ranks fixes by how much memory each one actually frees, so you try the effective ones first.
- A 2026 buying guide for the memory shortage, with a used-GPU inspection checklist.
- hwprofile.py — a dependency-free script that reads your RAM, VRAM and platform and prints exactly which models fit today.
Who it's for
- You are about to spend $400-$2,000 on a GPU and want to be right the first time.
- You already have a card and keep hitting "CUDA out of memory" without knowing which lever to pull.
- You are choosing between a used 3090, a new 4070 and a Mac — and every review compares gaming numbers, not tokens/sec.
- You run local models on a laptop or mini PC and need to know the real ceiling before you commit to a workflow.
Use cases
- Spec a machine for a specific model — start from the model you want, work back to the cheapest card that holds it.
- Audit a machine you already own: run hwprofile.py, then pull the winner for your tier from the matrix.
- Fix a slow setup — most "my GPU is slow" cases are a context-length or KV-cache setting, not the hardware.
- Buy used safely during the shortage, using the inspection checklist before you hand over cash.
Frequently asked questions
Is this the same material as the courses?
There is overlap. If you already have Pro or Lifetime, the hardware and model-fit material is covered across the courses and you do not need this — that is stated on the product page too. This exists for people who want the buying decision solved in one afternoon and have no interest in taking a course.
How current are the numbers?
The builds were assembled and measured between August and October 2025 and the model matrix is maintained against current releases. Component prices moved sharply through 2026 because of the memory shortage, so treat the parts lists as durable and the prices as a starting point — the kit says so explicitly and includes a section on buying in a shortage.
I use AMD / Apple Silicon, not NVIDIA. Is this useful?
Partly. Apple Silicon is covered directly, including the unified-memory maths and the sysctl that raises the GPU allocation limit. AMD is covered at the ROCm install level, but the measured speed figures are NVIDIA and Apple — there are no measured AMD numbers, and the kit does not pretend otherwise.
What format is it?
A zip of plain Markdown files, a Python script, and config files. No app, no login, no cloud service. Open it in any editor, keep it forever.
Do I need an account?
No. Buy it and the download link is emailed to you. If you do have an account, the purchase also appears in your library.
Are there refunds?
No — all sales are final, as stated in the terms. The full contents list above is deliberately specific so you can decide before buying rather than after.
Ready to get started?
One-time purchase · instant download · yours to keep.
Available soonAfter you buy
Purchases are linked to your account — sign in and head to your product library to download anytime. Bought without an account? Check your email for the download link and a one-click way to set a password.
← Back to all kits, tools & codebases