The 5-step recipe to train your own frontier model: open data, one arch, 1k GPUs

andrew_n_carr · x · 2026-09-20

A meme-style thread distilling the now-standard recipe for replicating a frontier model: pull the Nemotron and OLMo open datasets, format them into the jev shape with dsv4.1 flash, take the Needle 3 architecture and scale it up, secure 1k GPUs, and train. "Two years later, you've got your own."

The joke lands because it satirizes how formulaic open-source frontier-model replication has become: open datasets, a known architecture scaled up, and the only real ingredients left are compute and time.

Original post →

More from Fun

Fun channel →