CodeSync Blog
Ollama on a small machine
Ollama is the default provider. If it is not running, pnpm codesync providers says so and names the next free choice. Generate still writes the built-in template instead of failing the command.
This repository recommends qwen2.5-coder:7b when you have about 8 GB of RAM or about 6 GB of VRAM, and qwen2.5-coder:1.5b when you do not. Those figures are planning numbers from the model cards the README cites, not a benchmark from this site.
LM Studio and any OpenAI-compatible server, including llama.cpp's llama-server, are the other local options. A GGUF file in ~/.codesync/models is used only if a server or an optional node-llama-cpp install can actually run it. There is no fake completion.