The ladder
Model sizes, rung by rung
What each tier needs to train from scratch. Training is limited by GPU memory, not system RAM.
| Tier | Approx. size | GPU memory to train | Where it runs |
|---|---|---|---|
| T0 | ~20M parameters | ~1 GB | Any laptop; CPU smoke test done |
| First contact | ~34M parameters | 4.4 GB measured | 6 GB laptop GPU; run done July 2026 |
| T1 | ~200M parameters | ~5–6 GB | Borderline on a laptop; one 12–24 GB GPU |
| T2 | ~460M parameters | ~12 GB | One 24 GB GPU |
| T3 | ~1.5B parameters | ~30–40 GB | One 40–80 GB data-centre GPU |
| T4+ | 3B and up | 48 GB or more | A multi-GPU cluster |
Sizes are approximate. They assume the 152K-token Qwen2.5 vocabulary, whose embedding table dominates small models; a smaller custom vocabulary shrinks them. Training a 30B model from scratch is a cluster-scale job, never a single rented machine.