It's an encoder-only transformer that consumes past(blood glucose + carbs + insulin) and future(carbs + insulin) and predicts future blood glucose for the next 2 hours. Announced meals and boluses/basal are used to condition its predictions. The context size is variable (8 - 24 hours), and model can work in autoregressive mode to predict the next >2 hours. It also predicts time by looking at the context, but it never consumes time. The architecture is BERT-style: bidirectional attention with future BG masked. DILATE loss was used to fit the median line; pinball loss to fit the uncertainty bands. The two are "mixed" via Kendall-Gal. All blood glucose is in kovatchev risk space reparameterized to [40, 400] range. I have trained 4 model classes (nano, small, medium, large) and 3 variants for each (pretrained on simulator only, pretrained + finetuned on ohiot1dm, pretrained and finetuned on ohiot1dm + azt1d + shanghait1dm). The largest one has ~17 million parameters (16 heads across 16 layers). Pretraining for the largest model took ~48 hours. Finetuning took