Program-as-Weights: 0.6B Model Matches Qwen3-32B Prompting With 1/50 Memory
yuntiandeng · x · 2026-09-18
Yuntian Deng (Hugging Face) and collaborators published "Program-as-Weights (PAW)", a new paradigm that compiles fuzzy functions—log alerting, JSON repair, intent ranking—from natural-language specs into compact, locally-executable neural programs.
- A 4B compiler trained on FuzzyBench (a released 10M-example dataset) emits parameter-efficient adapters for a frozen lightweight interpreter.
- A 0.6B Qwen3 interpreter running PAW programs matches direct prompting of Qwen3-32B while using 1/50 the inference memory and running at 30 tokens/s on a MacBook M3.
- Key reframe: the foundation model becomes a tool builder—invoked once per function definition; subsequent calls are cheap, offline, and reproducible.
- Paper, code, and model weights are all public; hosted inference (150 ms remote) is available, with optional local inference for data control.
Related event: ProgramAsWeights Compiles Natural Language into Tiny Neural Programs(2 posts)→
More from Models
- Codex Reset Coupon Changes Usage Reset Cycle, User Flags Billing Quirk — algo_diver · 2026-09-19
- 28 hashes logged: crowd audits Gensyn's open-1b, first verifiable training run — benfielding · 2026-09-19
- Open-source auto-router prices every model per request, and shows warm cache beats cheaper models — airesearch12 · 2026-09-19
- Dev builds ErgoPulse, an open-source webcam posture app powered by DeepSeek V4.1 Flash — Arindam_1729 · 2026-09-19
- Bolt adds DeepSeek V4.1 Flash after week one: usage 10x DeepSeek V4 Pro — rohanpaul_ai · 2026-09-19
- Apple researchers propose probe guidance, cutting guidance cost for diffusion LMs with no extra forward pass — itsbautistam · 2026-09-19