Open research.
We study what it takes to run AI well, and how to make it cheaper.
How we work.
We measure on hardware people own. An Apple M5 Pro, an AMD Ryzen AI Max+ 395, an RTX 5070 Ti Laptop GPU. Each run is warmed up, repeated with the second pass reversed, and published with its raw data and its idle baseline.
Small models, one job each
Trace drafts tokens so a large model can accept several at once. Lattice keeps answers intact at 2 bits. Neither has beaten the stock model on our bar yet.
A model ships only when it beats the best open model on the same machine, per correct answer.
Open, and for profit
Every note states its prediction before the run, and keeps it when it is wrong. The code, the data and the weights of every model we release are open, under Apache-2.0. Spliced is a company, not a foundation; the research is open because measured claims are worth more when anyone can rerun them.
How we work
Measure
Repeated runs, held-out checks, raw data public.
Publish
Every note ships with its code and its limits.
Correct
When we are wrong, the note says so.
Found a mistake?
Open an issue. The note gets corrected, and says that it was.