Can a language model solve a harder problem by applying the same learned transformation a second time? NVIDIA’s Dory1.0-4B makes this question experimentally tractable. Its central block can process the same sequence at different depths while the model’s four billion parameters remain fixed. NVIDIA reports a substantial gain from the second pass on software repair and reasoning over long documents. Further passes yield smaller gains on average, and factual accuracy changes little on one knowledge test. This assessment examines the published weights, training design and evaluation results alongside measurements from people running the model on their own hardware. Experiments with llama.cpp and a Rust inference engine illustrate how precision, memory management and the choice of execution depth affect speed. These comparisons also show why the number of generated tokens, the time required to answer and the quality of the answer deserve separate measurements.
