The question
What is the minimum useful local inference node in 2026, measured in capability, watts, and dollars?
LocoFósil
Second-life GPUs, measured.
What is the minimum useful local inference node in 2026?
The floor after the floor
LocoBanco maps the consumer VRAM tiers. LocoFósil goes older: Kepler, Maxwell, Pascal, Turing, and the server cards that no longer get modern CUDA.
Nvidia first
One platform, documented properly. AMD and Intel results are welcome as community contributions once the Nvidia matrix is solid.
Measure everything
Tokens per second, prompt processing, tokens per joule, cost per million tokens, and whether the task actually succeeds.
What this is
Section titled “What this is”LocoFósil is a research project about hardware the modern AI stack has left behind. Not because the cards are interesting on their own, but because the floor is where most people actually are. A K1200 costs less than a new keyboard. An Optiplex costs less than a phone. If that combination can serve a useful 2B model at a usable speed and a sane wattage, that changes who can run local AI.
The shelf covers Kepler, Maxwell, and Pascal, from a 2 GB K620 to a 12 GB K6000. The inventory is on the hardware page.
The project has four tracks: a compatibility matrix, a benchmark suite, a low-power serving node, and a lightweight harness. The first two are measurement. The last two are proof that the measurements mean something.
Start here
Section titled “Start here”- Hardware: the cards, the compute capabilities, and what still runs.
- Benchmarks: what we measure and how.
- Low-power node: the Optiplex plus K1200 reference build.
- Lightweight harness: testing a 2B model with structure around it.
Dated notes and failed experiments live at LocoNotas.