---
id: 20260505-T0-09
title: "智能体工作流太烧钱？研究揭示小模型也能顶上"
title_en: "AgentFloor: Revealing How Small Models Can Handle Agentic Workloads"
url: https://ai.daily.yangsir.net/daily/20260505-T0-09
issue_date: 2026-05-05
publish_date: 2026-05-04T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2605.00334
---

# 智能体工作流太烧钱？研究揭示小模型也能顶上

在生产环境的智能体系统中，每次用户请求通常会触发多次模型调用，而其中大部分其实是简短、常规的结构化任务。针对这一成本痛点，研究人员提出并测试了AgentFloor框架，评估开源小模型在复杂工具调用场景下的表现上限。研究发现，在多步骤的Agent工作流中，许多基础且重复的节点完全可以由小模型接管，而不必每次都调用昂贵的旗舰大模型。企业可以利用这一结论优化路由策略，大幅削减API调用成本。

## English Version

**AgentFloor: Revealing How Small Models Can Handle Agentic Workloads**

Production agentic systems trigger multiple model calls per user request, mostly short and routine, raising operational costs. The AgentFloor framework evaluates how small open-weight models perform in these tool-use scenarios. Findings suggest routing routine tasks to smaller models can reduce reliance on expensive frontier models.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2605.00334)

**详情页**：https://ai.daily.yangsir.net/daily/20260505-T0-09

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*