---
title: "NariQ: Sub-3-Bit Quantization for... | SI Paper Desk"
url: https://thesimind.si/papers/2611.26157/
description: "NariQ compresses a 27B model to 3.0 or 2.5 bits per weight and, per the author, keeps it level with full precision on long agentic tasks by calibrating on..."
---

Paper of the Day · November 22, 2026

# NariQ: Sub-3-Bit Quantization for Long-Horizon Agentic Tasks

NariQ compresses a 27B model to 3.0 or 2.5 bits per weight and, per the author, keeps it level with full precision on long agentic tasks by calibrating on the model's own agent sessions.

Doyeob Kim · Submitted November 19, 2026 · arXiv 2611.26157v1 · cs.LG

[Read the paper on arXiv](https://arxiv.org/abs/2611.26157) · [The PDF](https://arxiv.org/pdf/2611.26157) · [The day's 12](https://thesimind.si/papers/day/2026-11-22/)

## What they found

The author, Doyeob Kim, says many low-bit quantization methods are calibrated on short web text and judged by perplexity. Quantization means storing a model's weights with fewer bits. The abstract says existing quantizations of Qwen3.8-27B at or below 3 bits per weight match BF16 on shorter tasks like AIME. But in published results they solve at most 72% as many tasks as BF16 on DeepSWE, a long-horizon agentic benchmark. NariQ is a trellis-coded post-training quantization method. It calibrates on tokens that the BF16 model generated in agent sessions. On Qwen3.8-27B, NariQ at 3.0 and 2.5 bits per weight stays level with BF16 on DeepSWE, Terminal Bench, AIME and RealWorldQA. The 2.0 bit checkpoint trails this performance.

## Why it matters

Agents that call tools over many steps and read 100K tokens or more of context are costly to run. Smaller models are easier to run on modest hardware. The abstract points out that short tests and perplexity can hide losses on long tasks. It also says the files are released as Qwen3.8-27B-NariQ-GGUF, with a llama.cpp fork for inference. The abstract gives no details on speed, memory use or how the method works beyond the calibration data.

## How it changes the race

If the result holds, strong agent models could run at under 3 bits per weight with little loss on long tasks. That could lower the hardware needed to run agents. It tests one model, so the wider effect is unclear.

### The abstract, by the authors

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.

Quoted from arXiv:2611.26157. The research and its words belong to its authors.

Desk score **75** out of 100

Why it was picked

- Agents (in the title)
- Reasoning (in the title)
- Evals
- Studies frontier models (Qwen3.8)
- Code or data released

Topics

- [Efficiency](https://thesimind.si/papers/?tag=efficiency)
- [Agents](https://thesimind.si/papers/?tag=agents)
- [Evals](https://thesimind.si/papers/?tag=evals)
- [Language models](https://thesimind.si/papers/?tag=language-models)
- [Benchmarks](https://thesimind.si/papers/?tag=benchmarks)
- [Training](https://thesimind.si/papers/?tag=training)

**arXiv ID:** 2611.26157v1
**Submitted:** November 19, 2026
**Fields:** cs.LG (Machine Learning)
**Level:** Intermediate
**Comments:** 7 pages, 1 figure, 3 tables

Fingerprint, SHA-256

`66e95a8efdb9ab8b3e7d7e03187336b70a867be40576d45b15dbab8abc9e69b3`

In SI MIND's journal since Nov 22, 2026, 02:01 UTC. It goes on Solana with the next hourly Proof of Mind seal.

### The exact text fingerprinted

{"arxiv":"2611.26157","version":1,"title":"NariQ: Sub-3-Bit Quantization for Long-Horizon Agentic Tasks","authors":["Doyeob Kim"],"day":"2026-11-22","rank":1,"score":75,"status":"written","summary":{"takeaway":"NariQ compresses a 27B model to 3.0 or 2.5 bits per weight and, per the author, keeps it level with full precision on long agentic tasks by calibrating on the model's own agent sessions.","found":"The author, Doyeob Kim, says many low-bit quantization methods are calibrated on short web text and judged by perplexity. Quantization means storing a model's weights with fewer bits. The abstract says existing quantizations of Qwen3.8-27B at or below 3 bits per weight match BF16 on shorter tasks like AIME. But in published results they solve at most 72% as many tasks as BF16 on DeepSWE, a long-horizon agentic benchmark. NariQ is a trellis-coded post-training quantization method. It calibrates on tokens that the BF16 model generated in agent sessions. On Qwen3.8-27B, NariQ at 3.0 and 2.5 bits per weight stays level with BF16 on DeepSWE, Terminal Bench, AIME and RealWorldQA. The 2.0 bit checkpoint trails this performance.","matters":"Agents that call tools over many steps and read 100K tokens or more of context are costly to run. Smaller models are easier to run on modest hardware. The abstract points out that short tests and perplexity can hide losses on long tasks. It also says the files are released as Qwen3.8-27B-NariQ-GGUF, with a llama.cpp fork for inference. The abstract gives no details on speed, memory use or how the method works beyond the calibration data.","race":"If the result holds, strong agent models could run at under 3 bits per weight with little loss on long tasks. That could lower the hardware needed to run agents. It tests one model, so the wider effect is unclear.","safety":"","level":"intermediate","tags":["efficiency","agents","evals","language-models","benchmarks","training"]}}

[Check it on Verify a thought](https://thesimind.si/proof-of-mind/)

## The authors

Doyeob Kim

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.

Related on the desk

## More papers on these topics

[Nov 18, 2026 · Number 2Nemotron 3.5 Bolt E4B 4GB: An Open and Efficient Small Language Model for Fast Conversational AgentsNVIDIA's authors post-trained a small open model from Gemma-4-E4B-it for fast local assistants. They report the top average accuracy on 21 benchmarks and faster decoding with a drafter model.EfficiencyLanguage modelsAgents](https://thesimind.si/papers/2611.20771/) · [Nov 16, 2026 · Number 4LuthorSQL: A One-Pass Text-to-SQL Model as System 1 for a Slower LLM AgentA small local 7B model can answer nearly half of database questions in one fast pass and hand the hard ones to a big SI agent, cutting API cost by 33.3% at the same accuracy.EfficiencyAgentsLanguage models](https://thesimind.si/papers/2611.17780/) · [Nov 22, 2026 · Number 2How Much Should an Agent Remember? Deployment Lessons from Self-Distilled Memory for LLM AgentsThe right amount of agent memory depends on the model: a strong model keeps gaining from more guidelines, while a weaker one did better with a small selected subset.AgentsMemoryEvals](https://thesimind.si/papers/2611.26361/)

[All 12 for Nov 22, 2026](https://thesimind.si/papers/day/2026-11-22/) · [Number 2](https://thesimind.si/papers/2611.26361/)
