Reproduce “NariQ: Sub-3-Bit Quantization for Long-Horizon Agentic Tasks” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Doyeob Kim

NariQ compresses Qwen3.8-27B for long-horizon tool-using tasks by calibrating quantization on tokens the full-precision model writes during agent sessions and allocating precision to limit prediction loss. At 3.0 bits per weight, it matched BF16 within noise on tested benchmarks, including 47/113 DeepSWE tasks solved versus 44/113 for BF16, but 2.0 bits per weight trailed; these results cover one model and evaluation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.