A research consortium comprising the University of Waterloo, Cornell University, and Harvard University has introduced Program-as-Weights, a novel framework designed to streamline local AI-assisted software development. Addressing the persistent challenges of cloud-dependent large language models, including data privacy risks, infrastructure costs, and version instability, the framework employs a compile-once, run-locally architecture. Instead of routing continuous queries to remote servers, the system leverages a large model to translate plain-language developer instructions into compact Low-Rank Adaptation modules. These lightweight adapters are subsequently integrated into miniature foundation models, enabling fully offline, device-resident execution for recurring programming tasks. Validated against FuzzyBench, a benchmark encompassing ten million fuzzy programming operations such as log parsing and JSON repair, the framework demonstrated superior performance relative to significantly larger systems. When paired with a quantized 430-megabyte interpreter, the system achieved a 73.78 percent accuracy rate, surpassing the 68.7 percent performance of Qwen3-32B, a model possessing more than fifty times the parameters. The architecture maintains high efficiency on standard consumer hardware, processing approximately thirty tokens per second on a MacBook M3. The research team has made the implementation and experimental code publicly available. By reframing foundation models as tool builders rather than per-query solvers, the framework advances a targeted paradigm where large language models handle compilation while smaller models manage runtime execution. The development aims to reduce computational overhead, eliminate cloud dependency, and provide developers with reliable, locally hosted automation for everyday coding workflows.