Menlo/Lucy

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:2BQuant:BF16Ctx Length:32kPublished:Jul 18, 2025License:apache-2.0Architecture:Transformer0.1K Open Weights Warm

Lucy is a compact 1.7 billion parameter model developed by Menlo Research, built on Qwen3-1.7B, and optimized for agentic web search and lightweight browsing. It features machine-generated task vectors and pure reinforcement learning for efficient thinking processes. Designed to run on mobile devices, including CPU-only configurations, Lucy excels at strong agentic search using MCP-enabled tools and basic browsing capabilities, demonstrating higher accuracy than DeepSeek-v3 on SimpleQA.

Loading preview...

Overview

Lucy is a compact 1.7 billion parameter model developed by Menlo Research, specifically designed for agentic web search and lightweight browsing. Built upon the Qwen3-1.7B architecture, Lucy inherits advanced research capabilities while being highly optimized for efficient execution on mobile devices, even with CPU-only setups. Its efficiency is achieved through the use of machine-generated task vectors that refine thinking processes and smooth reward functions across various categories, utilizing pure reinforcement learning without supervised fine-tuning.

Key Capabilities

  • Strong Agentic Search: Leverages MCP-enabled tools like Serper for robust Google Search integration.
  • Basic Browsing: Supports web browsing through tools such as Crawl4AI and Serper.
  • Mobile-Optimized: Engineered for lightweight operation, ensuring decent speed on CPU or mobile devices.
  • Focused Reasoning: Employs machine-generated task vectors to optimize its thinking for search-related tasks.

Good for

  • Developers building mobile applications requiring on-device agentic web search.
  • Integrating lightweight browsing capabilities into resource-constrained environments.
  • Use cases demanding efficient, focused reasoning for information retrieval tasks.

Evaluation Highlights

Lucy demonstrates impressive performance for its size, achieving higher accuracy than DeepSeek-v3 on the SimpleQA benchmark, using the same MCP methodology as other Menlo models like Jan-Nano.