Today, Small56.AI releases DumbMath Preview 1 0722. Despite its extremely small size of 665K parameters, this model outperforms a Atom-2.7M four times its size; although it is a preview version, it demonstrates arithmetic performance comparable to Qwen/Qwen2.5-0.5B.

Parameters: 665K
Training tokens: 0.52B
Training time: Under 2 hours on Colab T4

What the model cannot do

Conversation (no alphabet tokens exist)
Complex mathematics
Arithmetic outside its supported scope

arithmark

eval	model	score
ArithMark 2.0	mathbananamind-1.1	90.2%
	our model	62.5%
	qwen2.5-0.5B	63.0%
Parameters	mathbananamind-1.1	2900K
	our model	665K
	qwen2.5-0.5B	494000k

56m/DumbMath_Preview_1_0722

DumbMath Preview 1 0722

What the model cannot do

arithmark

56m/
DumbMath_Preview_1_0722