Date180725General agent

Lucy

A 1.7B model built for web search through tool calling, small enough to bring search-powered answers to phones and low-power devices.

1 min read

Overview

Parameters
1.7B
Base model
Qwen3-1.7B
Context
128k
Built for
Web search through tool calling
License
Apache 2.0

78.3%

SimpleQA accuracy (YaRN 130k)

Lucy is a 1.7B parameter model built on Qwen3-1.7B, optimized for web search through tool calling. The model has been trained to work effectively with search APIs like Serper, enabling web search capabilities in resource-constrained environments.

Try it in Jan

Performance

SimpleQA Benchmark

Lucy achieves competitive performance on SimpleQA despite its small size:

BenchmarkJan-nano-128k-MCP (YaRN 130k)Jan-nano-MCPJan-nano-MCP (YaRN 130k)Lucy (YaRN 130k)DeepSeek-V3-MCPChatGPT-4.5Baseline-MCPGemini-2.5-ProClaude-3.7-Sonneto3Grok-3o1
SimpleQA (Question answering accuracy)83.2%80.7%79.7%78.3%78.2%62.5%59.2%52.9%50%49.4%44.6%42.6%

The benchmark shows Lucy (1.7B) compared against models ranging from 4B to 600B+ parameters. While larger models generally perform better, Lucy demonstrates that effective web search integration can partially compensate for smaller model size.

Requirements

  • Memory:
    • Minimum: 4GB RAM (with Q4 quantization)
    • Recommended: 8GB RAM (with Q8 quantization)
  • Search API: Serper API key required for web search functionality
  • Hardware: Runs on CPU or GPU

Using Lucy

Quick Start

  1. Download Jan Desktop
  2. Download Lucy from the Hub
  3. Configure Serper MCP with your API key
  4. Start using web search through natural language

Demo

Lucy Demo

Deployment Options

Using vLLM:


vllm serve Menlo/Lucy-128k \
--host 0.0.0.0 \
--port 1234 \
--enable-auto-tool-choice \
--tool-call-parser hermes \
--rope-scaling '{"rope_type":"yarn","factor":3.2,"original_max_position_embeddings":40960}' \
--max-model-len 131072

Using llama.cpp:


llama-server model.gguf \
--host 0.0.0.0 \
--port 1234 \
--rope-scaling yarn \
--rope-scale 3.2 \
--yarn-orig-ctx 40960


Temperature: 0.7
Top-p: 0.9
Top-k: 20
Min-p: 0.0

What Lucy Does Well

  • Web Search Integration: Optimized to call search tools and process results
  • Small Footprint: 1.7B parameters means lower memory requirements
  • Tool Calling: Reliable function calling for search APIs

Limitations

  • Requires Internet: Web search functionality needs active connection
  • API Costs: Serper API has usage limits and costs
  • Context Processing: While supporting 128k context, performance may vary with very long inputs
  • General Knowledge: Limited by 1.7B parameter size for tasks beyond search

Downloads

Citation

Dao, A., Vu, D. B., Nguyen, A., & Buppodom, N. (2025). Lucy: edgerunning agentic web search on mobile with machine generated task vectors. arXiv:2508.00360.

arXiv

Talk to us