We built ProgramAsWeights (PAW) — https://programasweights.com

You describe a function in English — like "classify if this message is urgent" — and PAW compiles it into a tiny neural program (22 MB) that runs locally like a normal Python function. No API keys, no internet after compilation, deterministic output.

It's for tasks that are easy to describe but hard to code with rules: urgency triage, JSON repair, log filtering, tool routing for agents.

  pip install programasweights

  import programasweights as paw
  f = paw.compile_and_load("Classify if this is urgent or not.")
  f("Need your signature by EOD")  # "urgent"

Compilation takes a few seconds on our server. After that, everything runs on your machine. Each program is a LoRA adapter + text instructions that adapt a fixed pretrained interpreter (Qwen3 0.6B). The model itself is unchanged — all task behavior comes from the compiled program.

On our evaluation, this 0.6B interpreter with PAW reaches 73% accuracy. Prompting the same 0.6B directly gets 10%. Even prompting Qwen3 32B only gets 69%.

Also runs in the browser (GPT-2 124M, WebAssembly): https://programasweights.com/browser

You can also use it in your AI agents by copying the prompt here: https://programasweights.com/agents

Source: https://github.com/programasweights

Try it out: https://programasweights.com