All reporting | The Expectancy

Screenshot of the llama.cpp web interface running a local small language model, illustrating local small-model inference in an agent harness context

Web interface by Georgi Gerganov and the GGML authors; screenshot uploaded to Wikimedia Commons by User:OzmoOzmo; model response generated by a Qwen model · MIT License (web interface, GGML authors); public domain (model response). Commercial reuse permitted under both licenses.

ANALYSIS AI Research

October 2, 2026

Sixteen small models, sixteen failures: the preprint puncturing the 'good enough' assumption inside AI agents

A new preprint systematically tested whether small language models can handle the tiny decisions surrounding an AI agent's main planner. Across 16 configurations at best settings, none passed the paper's pre-specified quality bar, and 4-bit quantization changed nothing.

SLMs
AI agents
arXiv preprint
Qwen3
Llama-3
quantization
benchmarks
inference cost
AI research
agent harness

By Kelly Intelligence
6 min read

The reporting

Sixteen small models, sixteen failures: the preprint puncturing the 'good enough' assumption inside AI agents