AIR-LLM Broadcasts Weights for Memory-Free Edge Inference

AIR-LLM attacks the memory-movement bottleneck with an unusually radical hardware-software split, but its headline gains remain simulation- and testbed-backed rather than proof of phone-ready deployment.

• –One broadcast can serve many users, making the architecture increasingly attractive as edge-device density rises.
• –RF-domain GEMV keeps prompts and computation at the edge while relocating weight storage to the network.
• –Channel quality, path loss, calibration, thermal noise, and mixer distortion materially affect accuracy.
• –The paper evaluates ray-traced Paris and Munich channels plus a measured RF mixer, not a complete over-the-air commercial system.
• –For developers, this points toward cellular infrastructure as an inference substrate—not merely a connectivity layer.