Overnight runs of 2026-11-12: the 18M model with gated attention and QK-norm (1.393 against 1.419), with the circuit search (1.431), the 0.1B model at step 3,000, the 8,000-step wide fine-tune, and their result files
Research checkpoints of the 17M experiments (checkpoints/17m), all result files (results/), and an updated card: both memory skills at 17M, equal-compute results, a correction to the swarm-on-text claim
