Waterloo + Cornell + Harvard · 2026-07-02 · notable
Program-as-Weights — compile English into LoRA adapters for a local 0.6B model
Program-as-Weights (PAW) is a compiler that turns natural-language specs like 'repair broken JSON' into a LoRA adapter for a frozen 0.6B model. Runs at 30 tok/s on a MacBook M3 and matches Qwen3-32B prompting with roughly 50x less memory.
Compile English task specs into LoRA-shaped neural programs that run locally and match a 32B model.
What is it?
Program-as-Weights (PAW) is a new paradigm for the fuzzy tasks that resist rule-based code — log triage, JSON repair, email classification, key extraction. Instead of hitting an LLM API on every input, PAW compiles a natural-language specification once into a small neural artifact that runs on your own machine.
How does it work?
The system splits work in two stages. A 4B compiler reads the spec plus a handful of examples and emits both a paraphrased pseudo-program and a LoRA adapter. A frozen 0.6B interpreter then loads that adapter and executes the compiled program on user input. Heavy compilation runs once; every call after that is cheap and local.
Why does it matter?
Program-as-Weights matches Qwen3-32B prompt-based quality on the paper's fuzzy-function benchmark while using roughly 50x less inference memory. The interpreter reaches about 30 tokens per second on a MacBook M3 — no API bill, no data leaving the device, and the compiled program is a 22 MB file you can ship with your app.
Who is it for?
engineers replacing hand-written parsers or per-request LLM calls with a small, offline model.
Try it
programasweights.com — live demo + Python SDK