Two hours is a price. The page never says on what.

64M parameters, from scratch, two hours. The first two entries close cleanly. The third one needs a machine name to settle, and the description does not give one.

$ git clone minimind

github.com/jingyaogong/minimind

Claim ledgerEntries from the description only

entry

settles

64M parametersA count anyone can check by loading the config. It also sets the ceiling on what the model can do.

closed

from scratchNo pretrained weights. This is the entry that makes the other two mean something.

closed

in just 2hA wall-clock cost with no hardware attached. Two hours on one consumer card and two hours on a rented cluster are different products.

open
no machine named

the entry a reader has to reproduce2hdenominator missing

Small is the point

64M is a size you can hold in your head and retrain after a bad idea. That makes it a teaching artifact rather than a competitor to anything on a leaderboard.

From scratch is the load bearing phrase

Fine-tuning a small model in two hours is ordinary. Starting from random weights and stopping at two hours is the claim worth reading twice.

What is not on this page

No GPU, no dataset, no token count, no loss curve, no eval number. The trending description names none of it, so neither does this layout.

Concept mock, not affiliated with the project. The GitHub Trending description on 31 August 2026 reads "Train a 64M-parameter LLM from scratch in just 2h!" The ledger lists those three claims and marks the third one open because no hardware appears in that sentence. No benchmark, dataset, or training curve is represented here.