Model 001Released July 2026

TinyRamp

0.5B parameters2 primary tasksPortable weights

Our first small model. Trained with Distillery. Evaluated against the alternatives.

One model · two finance jobs

A finance generalist small enough to carry.

TinyRamp handles transaction review and variance analysis with the same weights. No hidden specialist router. No second checkpoint behind the demo.

Read the experiment See how it was trained

Amazon Nova Pro teacherNova

TinyRamp · Qwen2.5-0.5B student0.5B

sequence.v1rejection sampling

Release evaluation

Evaluated, within the experiment.

On the sealed 40-example held-out set, TinyRamp reaches 55% schema validity against the teacher's 57.5% and matches the teacher's 25% decision-field accuracy at roughly a thousandth the size. All claims remain bounded to the synthetic finance benchmark; serving economics are projected, not measured.

Schema validity (40 held-out) 55% vs teacher 57.5%

Decision-field accuracy 25%, teacher parity

Systems behavior Measured

Economics Projected

Built with Distillery

Bring a dataset. Keep the model.

Open Distillery