Compile English specs into 22 MB neural functions that run locally
Details
- External ID
- 47780141
- Source
- HN
- Company
- —
- Product
- —
- Website domain
- —
- Launched
- April 15, 2026
- Cohort
- —
- Upvotes
- 11
- Upvotes percentile
- 0.62146529562982
- Tags
- —
- Fetched at
- Sept. 7, 2026, 9:26 p.m.
- Updated at
- Sept. 7, 2026, 9:26 p.m.
Description
We built ProgramAsWeights (PAW) — https://programasweights.comYou 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/browserYou can also use it in your AI agents by copying the prompt here: https://programasweights.com/agentsSource: https://github.com/programasweightsTry it out: https://programasweights.com
Enrichment
- Theme
- lightweight and on-device AI runtimes
- Vertical
- Horizontal
- Function
- Model & infra
- Audience
- Developer
- AI stance
- AI-native
- Project type
- Commercial product
- Normalized one-liner
- compile specifications into local neural functions
- Manually corrected
- False
Could you build this?
No Compiling natural language specifications into ultra-small, deterministic 22 MB executable neural network weights requires novel ML research in weight generation, meta-learning, and model distillation.
What it would actually take: This architecture likely utilizes a meta-model or hypernetwork trained to emit model weights directly from text specifications, or an automated pipeline combining synthetic dataset generation, architecture search, and aggressive quantization/distillation into tiny sub-20M parameter architectures (e.g., heavily quantized MobileNet/tiny Transformer backbones). Achieving deterministic inference within a tiny binary without external dependencies requires deep ML systems engineering, custom C++/Rust inference runtimes (like ONNX or llama.cpp bindings), and massive GPU compute for training the compiler model.
Discussion
No comments on this launch.
Competitors
Other products that read as similar to this one — 16 launches clear the similarity bar, closest 8 shown.
Attention rank: #8 of 17 (itself plus its competitors, highest first — normalized so YC and Product Hunt are compared fairly).
Launched 157 days after the earliest competitor.
- Linear RNN/Reservoir hybrid generative model, one C file (no deps.) · hn · 2026-04-09 · 7 upvotes · similarity 0.42
- tiny-classifiers · github · 2026-09-27 · 68 upvotes · similarity 0.39
- Runprompt · hn · 2025-11-27 · 134 upvotes · similarity 0.36
- GlyphLang · hn · 2026-01-10 · 44 upvotes · similarity 0.36
- Sigil · hn · 2026-07-24 · 5 upvotes · similarity 0.34
- I embedded 685M public texts in 32 minutes (on 8x A100, Rust, TensorRT) · hn · 2026-06-04 · 7 upvotes · similarity 0.34
- Built a tiny interpreter from scratch in C to understand how they work · hn · 2025-11-12 · 5 upvotes · similarity 0.33
- I Made a Programming Language with Python Syntax, zero-copy and C-Speed · hn · 2026-02-18 · 10 upvotes · similarity 0.33
Other launches for this product
- No other launches for this product.
Same idea, different domain
Nobody's really built a model & infra tool for Fintech yet.