Why now
Owning AI just became possible.
Two years ago it took a research lab. Today it takes a box and a method.
The numbers
Four figures, every one sourced.
- How far open models trail the best closed onesEpoch AI, open and closed models on the Epoch Capabilities Index, 29 May 2026
- 4 months
- Points above GPT-4 for tuned open models on narrow tasks, on average across 31Predibase, LoRA Land, arXiv 2405.00732, 2024
- +10
- Memory in a desk machine that now fine-tunes models of up to 70B parametersNVIDIA, DGX Spark, 13 October 2025
- 128 GB
- When the EU AI Act’s high-risk duties apply (Annex III). Annex I follows on 2 August 2028.Regulation (EU) 2026/1744, Official Journal, 24 July 2026
- 2 Dec 2027
- 01
Open models caught up.
Open models trail the best closed ones by about four months. Tuned on one narrow task, they beat far larger models on that task.
- 02
The machine fits under a desk.
A desk machine with 128 GB of memory now fine-tunes models of up to 70B parameters. A first team needs no data centre.
- 03
The rules point home.
Data residency laws, the EU AI Act and export rules keep moving data and models closer to the company that owns them.
Start
The timing works. Does the task?
Two minutes, an honest answer, no email needed.



