Program-as-Weights: 0.6B interpreter matches Qwen3-32B prompting with 1/50 the memory
yuntiandeng · x · 2026-09-16
Yuntian Deng's team released the open-source paper "Program-as-Weights: A Programming Paradigm for Fuzzy Functions."
- Idea: fuzzy tasks (log alerting, JSON repair, intent-based ranking) are usually outsourced to LLM APIs at the cost of locality, reproducibility, and price. PAW compiles a natural-language spec into a compact, locally executable neural artifact.
- Method: a 4B compiler trained on FuzzyBench (a released 10M-example dataset) emits parameter-efficient adapters for a frozen lightweight interpreter.
- Results: a 0.6B Qwen3 interpreter running PAW programs matches direct prompting of Qwen3-32B, using 1/50 the inference memory at 30 tokens/s on a MacBook M3.
- Paradigm shift: the foundation model becomes a tool builder—invoked once per function definition to produce a small reusable artifact whose subsequent calls are cheap and offline.
Paper and code are open source.
More from Infra
- SentencePiece Lite ships: 50KB binary, 20-30x faster tokenization for edge devices — heiga_zen · 2026-09-16
- Leaked 10,000-word Huawei memo: become the NVIDIA for any LLM, pivot around Ascend 950 — pstAsiatech · 2026-09-16
- Running Qwen 27B and DeepSeek v4 Flash together on one heterogeneous machine — samsja19 · 2026-09-16
- JPMorgan sees 25M+ GPU/ASIC shipments by 2028, ASICs dominate — a 'narrative violation' — bookwormengr · 2026-09-16
- iamtrask: The Endgame Is a Trust Web of Personal LLM Servers, Not One AGI — iamtrask · 2026-09-16
- Intel bets on CPU-side KV Cache offloading and QAT hardware compression for agent inference — 量子位 · 2026-09-16