Satyrn’s own engine (Apple Silicon)

Besides Open Interpreter on Ollama, this project contains satyrn-codegen, a small coding CLI on top of satyrn-agent. That engine runs the dense Qwen3.8-27B model with DFlash speculative decoding, where a small drafter model proposes tokens that the large model checks in parallel.

Important

satyrn-agent’s dense model path runs on Apple’s Metal only. You need an Apple Silicon Mac with macOS 15 or newer and 32 GB or more of unified memory. On Linux and Windows, use Open Interpreter as described in Set up; the same base model is available there as qwen3.8:27b (see Choose a model).

This path has not been tested on a Mac by this project yet.

Setup

  1. Clone satyrn-agent next to this project, so that both folders share a parent folder.

  2. Download the pinned model pair as described in satyrn-agent’s README (section Models): Qwen3.8-27B-Q4_K_M.gguf (the model) and dflash-Qwen3.8-27B-Q4_0.gguf (the drafter).

  3. Build the patched libds4 in satyrn-agent: ds4-check --model qwen35 --build. The library goes to ~/.ds4py/lib/libds4.dylib.

  4. Install the CLI with pipx and run it:

make install-codegen   # builds satyrn-agent and satyrn-codegen wheels, installs them with pipx
export SATYRN_MODEL=~/models/Qwen3.8-27B-Q4_K_M.gguf
export SATYRN_DRAFTER=~/models/dflash-Qwen3.8-27B-Q4_0.gguf
satyrn-codegen "create a program that takes a number and outputs the square root of it"
satyrn-codegen -i  # conversation; /exit leaves

The model asks for tools (list, read, write and edit files, run commands), and you approve every change unless you pass -y. Other options: -C DIR (folder to work in), --dflash auto|on|off, --think none…max, --no-commands, --show-thinking. satyrn-codegen --help lists all of them.

The same square-root test exists for this path: uv run pytest -m live tests/test_live_sqrt.py.