Small model predicts T1D blood sugar zero-shot

A developer trained a 31,251-parameter encoder-only transformer on synthetic T1DM patient simulator data to predict blood glucose 2 hours ahead, then tested it zero-shot on real CGM traces. It supports counterfactual reasoning and 8-hour nocturnal predictions, trained in under 60 minutes on an NVIDIA DGX Spark, with the Android app using the ExecuTorch backend.

• Ling 3.1 Flash 560B MoE deployed across four DGX Sparks via llama.cpp2026-10-18
• Hugging Face releases tokenizers v1 with big speed gains2026-09-21