Sketchformer is a novel transformer-based representation encoding free-hand sketches input in a vector form, .e. as a sequence of strokes. Sketchformer effectively addresses tasks: sketch classification, sketch based retrieval (SBIR), and the reconstruction and interpolation sketches. We report several variants exploring continuous tokenized input representations, and contrast performance. Our learned embedding, driven by a learning tokenization scheme, yields state of the performance in classification and image retrieval tasks, compared against baseline representations driven by sequence to sequence architectures: SketchRNN and . We show that sketch reconstruction and interpolation improved significantly by the Sketchformer embedding complex sketches with longer stroke sequences.