Programs as Weights
Program-as-Weights matches Qwen3-32B using 50x less memory, runs at 30 tok/s on a MacBook M3
A new paradigm compiles natural-language function specifications into compact neural artifacts using a 4B-parameter compiler and 0.6B-parameter interpreter. The approach achieves performance parity with Qwen3-32B while using 50x less inference memory, running at 30 tokens per second on a consumer MacBook M3 — a fundamental step toward running frontier-level intelligence on edge devices.
