ProgramAsWeights: describe an AI function in English, compile it once, and run it locally on CPU
ProgramAsWeights (PAW) is a project from the University of Waterloo that allows users to describe an AI function in English, compile it once, and run it locally on a CPU. This initiative addresses the growing interest in open-source and locally executable AI alternatives. Resources available include a Python SDK, compiler weights (paw-4b-qwen3-0.6b), a research paper (arxiv.org/abs/2607.02512), and a playground for experimentation.
Time & source
Times shown in UTC
Display time zone: UTC
Local time zone unavailable; showing UTC.
PublishedOffset at this time: UTC+0Sep 18, 2026, 20:06 UTC
IngestedOffset at this time: UTC+0Sep 19, 2026, 13:00 UTC
- Published
- Sep 18, 2026, 20:06
- Ingested
- Sep 19, 2026, 13:00
- Source type
- Dev community
- Tier
- Community
- Source status
- Healthy
Tier is a per-source editorial setting, not a per-item score.
With the recent interest in Jev, especially in open-source and locally executable alternatives, I wanted to share a related project we've been building at the University of Waterloo: ProgramAsWeights (PAW).
The idea is simple: describe a function in English, compile it once, then call it from Python. The resulting function runs on your own CPU, without sending each input to an API.
https://preview.redd.it/xlsfp5rl6cqh1.png?width=1800&format=png&auto=webp&s=fb05ea259fdbf53c965a8373a3c1fc9688347cab
import programasweights as paw fn = paw.compile_and_load("Classify urgent emails") fn("Need this today") # "urgent" (runs locally)
We trained a larger model to generate a LoRA adapter from an English function description. That adapter specializes a shared Qwen3 0.6B interpreter for the task. You can keep multiple compiled functions and reuse the same base model.
The SDK uses our hosted compiler by default. Once the compiled program and base model are downloaded, inference runs locally and works offline. The compiler weights are also public for people who want to run compilation themselves.
I've used this to build a course website helper with ~30 small neural programs connected by ordinary decision-tree code. One function decides which answerer should handle a question, and the surrounding code controls what happens next. Here's my course website helper in case you are curious: https://yuntiandeng.com/teaching/spring2026/cs486-introduction-to-artificial-intelligence/#ask
Some of you may remember an earlier version posted here (https://www.reddit.com/r/LocalLLaMA/comments/1sm9fmw/compile_english_function_descriptions_into_22mb/ ). At the time, people asked for the compiler itself to be released. Its weights are now public, along with the paper explaining how it works.
Python SDK: https://github.com/programasweights/programasweights-python Compiler weights: https://huggingface.co/programasweights/paw-4b-qwen3-0.6b Paper: https://arxiv.org/abs/2607.02512