Program-as-Weights: 0.6B model matches Qwen3-32B prompting with 1/50th memory, runs locally
yuntiandeng · x · 2026-09-21
Yuntian Deng's team introduces Program-as-Weights (PAW), a paradigm that compiles fuzzy functions from natural-language specs into compact, locally executable neural artifacts. The paper and the 10M-example FuzzyBench dataset are public.
How it works:
- A 4B compiler turns English function descriptions into LoRA adapters for a frozen 0.6B Qwen3 interpreter—no per-task finetuning
- The 0.6B interpreter running PAW programs matches direct prompting of Qwen3-32B while using roughly 1/50th the inference memory
- Runs at 30 tokens/s on a MacBook M3; subsequent calls are cheap and offline
Target tasks include log alerting, repairing malformed JSON, and intent-based search ranking. PAW reframes foundation models from per-input problem solvers into tool builders: invoked once per function definition, producing a small reusable artifact. Evaluation uses task specifications unseen during training.
More from Infra
- How do you test LLM provider failure in production? A Reddit discussion — Rama_Surasani_ · 2026-09-21
- Run Flux 2 Dev (30B) locally with block offloading, mix models for T2I and editing — Altruistic_Heat_9531 · 2026-09-21
- Jensen Huang stands by $3-4T AI infrastructure market forecast — emmanuelvivier · 2026-09-21
- Free Zoom meetup: disaggregated speculative decoding on d-Matrix chips plus inference engine tuning — cfregly · 2026-09-21
- MiMo near-SOTA on DeepSWE with just ~$2.6M RL run: will data cost more than training? — my_cat_can_code · 2026-09-21
- Personal AI Agents Are the Biggest Driver of the Sudden NAND Demand Surge — zephyr_z9 · 2026-09-21