GitHub radar
Run a tool-calling AI model on your own laptop
This is a compressed version of the Qwen3.8-27B model that can call tools and runs on a laptop or Mac instead of a big server.
It's a shrunk-down version of a large language model, Qwen3.8-27B. It understands requests and can call outside tools for an AI agent.
When it helps
Good for a local agent doing simple, repetitive tasks on your own machine. Not a fit if you need heavy math or deep reasoning.
Pros
- Only 7.89 GB instead of 54 GB — fits a regular laptop
- Calls tools better than the original: 88 of 120 tasks vs 84
- Keeps 96% of the original's performance on average across 9 benchmarks
- Can add image understanding with a separate vision file
Cons
- Much weaker at competition math: 79.2 vs 96.7 on AIME 2025
- Struggles with letter-level puzzles — palindromes, vowel counts, alphabetical order
- About a fifth of parallel tool-call replies have small formatting errors
How to set it up — step by step
- 1Check your computer (laptop or Mac) has 16 GB+ RAM — the model file is about 8 GB.
- 2Open your AI agent (Claude Code, Codex) in your project folder.
- 3Send the agent the text from the block below.
- 4Have the agent download the model, run it with llama.cpp, and register it as a tool-calling model.
- 5Test it: ask the agent to give the model a sample task that calls a tool.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up this compressed Qwen3.8-27B build by Underdog locally for me: https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0 — run it through llama.cpp, hook it up as a tool-calling model for an agent, and show me how to give it a task.
A regular laptop or Mac with 16 GB of RAM or more — the quantized file is about 8 GB.
Open on Hugging Face▌ More finds