paw-4b-qwen3-0.6b — ProgramAsWeights "Standard" compiler
This is the Standard compiler from ProgramAsWeights (PAW). Given a natural-language spec (a description of a function), it emits a tiny per-task program — a LoRA adapter — that then runs locally on the Qwen3-0.6B interpreter.
It is the model invoked by paw.compile(spec, compiler="paw-4b-qwen3-0.6b") and powers programs on https://programasweights.com.
- Compiler base model:
Qwen/Qwen3-4B-Instruct-2507 - Target interpreter:
Qwen/Qwen3-0.6B - Snapshot:
20260407(see git tag20260407)
Contents
compiler/— a finetuned Qwen3-4B-Instruct-2507 causal LM (the compiler).lora_mapper.pt— the mapper head (trunk + coefficient head + learnable LoRA basis matrices) that turns the compiler's hidden states into a LoRA program.meta.json—lora_rank=64,lora_alpha=16,lora_num_bases=64,prefix_steps=64, target modules[q,k,v,o,gate,up,down]_proj.
How it works
- The 4B compiler generates a short "pseudo-program" (a task description plus a few I/O examples) from the spec.
- The text
chat_template(spec) + pseudo-program + 64 prefix tokensis run through the compiler; the mapper reads the 64 prefix hidden states and emits per-layer LoRAA/Bmatrices as a learned mixture of basis matrices. - The resulting LoRA (about 22 MB) is the program. It loads onto Qwen3-0.6B and runs locally and offline (about 100 ms/call).
Status
- Inference/runtime SDK (load + run a compiled program locally): https://github.com/programasweights/programasweights-python
- The cleaned compile/runtime code and the arXiv preprint ("Program-as-Weights: A Programming Paradigm for Fuzzy Functions", AIware 2026) will be public by Jul 6, 2026. An uncleaned reference snapshot is at https://anonymous.4open.science/r/programasweights
- Live demo + program hub: https://programasweights.com
- Paper: arxiv.org/abs/2607.02512