Little Coder
Little Coder Overview
Little Coder (little-coder, or lc) is an offline coding agent built on top of the pi coding agent framework, tuned for small local models. On the official Terminal-Bench 2.0 leaderboard, little-coder + Qwen3.6-35B-A3B achieved 24.6% ± 3.2 accuracy — ranking above Gemini CLI + Gemini 2.5 Pro's official submission, all running on an 8 GB laptop GPU with the wifi off.
- Pi Framework Based: Built on top of the pi coding agent framework, not Raspberry Pi — pi is the name of the underlying agent framework.
- Small Model Optimized: Specifically tuned for small local LLMs like Qwen3.6-35B-A3B (35B MoE, 3B active parameters), which can run on 8 GB laptop GPUs with appropriate quantization via llama.cpp.
- Terminal-Bench Leaderboard: Achieved 24.6% ± 3.2 accuracy on Terminal-Bench 2.0, ranking above Gemini CLI + Gemini 2.5 Pro.
- Offline Operation: Runs entirely offline with no internet connection required.
- Official Install: Install with
npm install -g little-coder— the official installation method. - Local LLaMA.cpp Integration: Connects to your locally running llama.cpp server (OpenAI-compatible API) for model serving.
- Project File Management: Reads, edits, and runs commands against your project working tree with precise context tracking.
- Open Source: MIT licensed, developed transparently on GitHub.