garagelmthe garage intelligence lab · real models, cheap hardware, honest numbers
home · models · notes · competitions · learn · team
$ purpose
Frontier model development is a recipe, not magic. garagelm exists to prove the whole recipe runs on hardware anyone can buy. We train the models, measure them honestly, release the weights, and publish every number.
$ team
The lab is small on purpose: one human, one Mac, and a mascot.
AI engineer at American Express and Georgia Tech M.S. (AI). Trains small LLMs on consumer hardware, including a 232M model to GPT-2-class benchmark scores in 127 hours on a Mac mini. Trains the models, breaks the toys, writes the notes.
[02] you, possibly ● open · call for help
The lab is looking for collaborators. Bring a fair comparison, a replication, a good negative result, or compute, data pipelines, eval harnesses, and write-up reviews. Issues and pull requests all read.
$ hardware
[01] apple m4 pro ● on duty
48GB unified memory, no CUDA, no complaints. Does the actual work: ~127 hours per flagship, resumable every 500 steps. Uptime is its whole personality.
[02] hyperscaler gpu (placeholder · the garage has room)
Reserved for the day a sponsored H100 (or a very generous cloud credit) shows up. It will be held to the same rule as the Mac: fair comparisons, gates written before results.
$ outputs
The flagship base model: matches Pythia-160M at 300× fewer tokens (600× at its 0.5B checkpoint). weights ↗
The SmolTalk SFT variant: chat formatting with zero benchmark regression. weights ↗
188 parameters of pure transparency. Lives in learn, gets overfit daily, has never once hidden a number.
reach the lab: trevino293@gmail.com · github issues preferred