Preprint runs quantum replacements for language-model blocks on IBM hardware

Researchers report replacing eight consecutive transformer blocks in Qwen3-30B-A3B-Instruct-2507 with a quantum module containing 145,281 trainable parameters, displacing 16.33% of the model's parameters. The preprint describes splitting the model's hidden space into parallel quantum circuits linked by classical layers. The researchers trained with exact classical simulation, then executed the quantum portions on a 156-qubit IBM Heron processor. The experiments demonstrate hardware execution, but do not establish an advantage from trained, entangling quantum circuits: several simplified controls performed better on held-out perplexity, and a classical network with a matched parameter count remained a few percent ahead.