Skip to content

Releases

Models (models-2026-11-16)

Choose a tag to compare

@greydanus greydanus released this 17 Nov 03:22

hand.npz: the handwriting model, trained by this repository on seven sources (6.1M parameters, 208 wide, 1,192 hands, step 5,625 of --profile 8xa100 --n_embd 208 --steps 45000), 5.8 MB.
sketch.npz: the drawing model, trained by this repository on ten sources and told SigLIP 2 vectors (54M parameters, step 16,500 of 18,000), 50 MB.
siglip-text.npz: the text half of SigLIP 2 (google/siglip2-so400m-patch14-224, Apache 2.0), packed, 375 MB; it turns typed words into the vector the drawing model is told.

Both models are packed by tools/pack.py at 8 bits a weight, which costs 0.001 nats a token for handwriting and 0.003 for drawings against the weights as trained, and run with inference/pen.py:
python inference/pen.py write "The quick brown fox" --writer sam
python inference/pen.py draw "a cat|a sailboat"

hand.pt and sketch.pt: the same two models as trained, before packing, for the PyTorch code (24 MB and 216 MB). hand.pt names its writers by the ids hand.npz uses, which are the ids of the earlier release.
python pengpt.py sample --checkpoint hand.pt --writer sam --greedy --text "The quick brown fox"

Since models-2026-11-12: the handwriting model is a little wider and trained a little longer (0.711 nats a pen move on held-out words by known writers, from 0.716), and both models are packed at 8 bits in place of 7. The drawing model is the same model as before; sketch.pt is unchanged.

inksight-weights.npz and inksight-spm.model: InkSight (Derendering/InkSight-Small-p, Apache 2.0), its weights and tokenizer as inference/hand.py reads them, the weights stored as float16, 0.7 GB; hand.py downloads them into inksight/ the first time it traces a photo. Unchanged.