NariQ: Sub-3-Bit Quantization for Long-Horizon Agentic Tasks
Organizations: Nari Labs
Abstract
Long-horizon agentic tasks require a model to call tools over many steps at contexts of 100K tokens or more, yet many low-bit quantization methods are calibrated on short web text and judged on perplexity. This mismatch is costly: existing quantizations of Qwen3.8-27B at or below 3 bits per weight match BF16 on shorter tasks like AIME, yet in published results solve at most 72% as many tasks as BF16 on DeepSWE, a long-horizon agentic benchmark. We propose NariQ, a trellis-coded post-training quantization method that calibrates on tokens generated by the BF16 model in agent sessions. On Qwen3.8-27B, NariQ at 3.0 and 2.5 bits per weight stays level with BF16 on two long-horizon agentic benchmarks, DeepSWE and Terminal Bench, as well as on AIME and RealWorldQA. Our checkpoint at 2.0 bits per weight trails this performance. To our knowledge, these are the first quantizations at or below 3 bits per weight shown to match their full-precision version on long-horizon agentic tasks. We release the files as Qwen3.8-27B-NariQ-GGUF (https://huggingface.co/nari-labs/Qwen3.8-27B-NariQ-GGUF), with inference available through our llama.cpp fork at https://github.com/nari-labs/nari-llama.cpp.
Figures & tables
| DeepSWE v1.1 | Terminal Bench 2.1 | ||||
| File | bpw | Solved | % BF16 | Score | % BF16 |
| Our harness | |||||
| BF16 | 16 | 44 | 100 | 62.6 | 100 |
| Ours 3.0 | 3.00 | 47 | 107 | 63.8 | 102 |
| Ours 2.5 | 2.50 | 46 | 105 | 60.2 | 96 |
| Ours 2.0 | 2.00 | 33 | 75 | 52.4 | 84 |
| File | AIME | RealWorldQA | MGSM |
|---|---|---|---|
| BF16 | 97.7 | 82.4 | 94.0 |
| Ours 3.0 | 98.3 | 82.4 | 93.1 |
| Ours 2.5 | 96.5 | 81.3 | 91.8 |
| Ours 2.0 | 86.0 | 80.9 | 90.6 |
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| File | bpw | Agent | Reserve | General | Math |
|---|---|---|---|---|---|
| Ours 2.0 | 2.00 | 0.0284 | 0.0270 | 0.0841 | 0.0829 |
| Unsloth UD-IQ1_M | 2.00 | 0.1823 | 0.2071 | 0.2931 | 0.3105 |
| Unsloth UD-IQ2_XXS | 2.14 | 0.0732 | 0.0851 | 0.1115 | 0.1024 |
| EXL3 SC_1.40 | 2.18 | 0.3135 | 0.3614 | 0.5536 | 0.5809 |
| Unsloth UD-IQ2_XS | 2.27 | 0.0593 | 0.0686 | 0.0899 | 0.0819 |
| EXL3 SC_1.60 | 2.36 | 0.2415 | 0.2813 | 0.4181 | 0.4310 |
LLM Quantization
Large language model quantization reduces numerical precision to lower memory use and accelerate training or inference.
See all 448 papers