What matters in AI.

Subscribe

NariQ cuts a model to 2.5 bits and keeps its agent results

NariQ stays level with the full model at 3.0 and 2.5 bits, where other methods did at most 72% as many long agent tasks.

Researchers gave NariQ, a procedure to cut the size of a language model. It uses less bits for each weight, which is a number that the model keeps. This is quantization, and it cuts memory use. Other methods were tuned on short internet text, and they had lower results on long agent tasks. NariQ tunes on text that the full model writes in agent sessions.

How it works

NariQ changes a complete model and does not train it again. It tunes the step that cuts the size on tokens that the full model, BF16, wrote in agent sessions.

The numbers

NariQ at 3.0 and 2.5 bits for each weight stays level with the full model on 2 long agent benchmarks and 2 other tests. The 2.0 bit file is lower. In the known results for other methods at 3 bits or less, they did at most 72% as many tasks on DeepSWE.

What is not known

The authors write that these are the first files at or below 3 bits that agree with the full model on long agent tasks. The tests use one model, Qwen3.8-27B. The text gives no data on speed or memory.

Insight: Short tests hid the loss of small models

Other files agree with the full model on the short AIME test, but not on long tasks. Nvidia told that token use is 2 times more each month. Agents with a lower cost possibly need tests that agree with long agent work, thus that test shows which cost cut works. Tell vendors to show long task results.

Sources

  1. NariQ: Sub-3-Bit Quantization for Long-Horizon Agentic Tasksarxiv.org
AI MATTER · NEWS · AI MATTER · NEWS ·20 NOV2026

Posted

Tags

Learn the terms in this story

Guides for this story

More in Research

All Research news