autodecode v0.1.0
Ornith-9B on Apple Silicon — first public release
Highlights
- Karpathy-style autonomous inference loop (
prepare.py+optimize.py+program.md) - 21.4 decode tok/s on M4 32GB (vs ~10 Ollama, ~17.5 MLX baseline)
- 22 documented experiments in
results/experiments.tsv - 3-pass stable benchmark harness with noise margin
- MLX server script for OpenCode / Hermes
- Full results write-up in
RESULTS.md
Hardware validated
- MacBook Pro, Apple M4, 32 GB RAM
- Ornith-1.0-9B MLX 4-bit (~4.7 GB)
Breaking / known
- Model weights not included — run
mlx_lm.convertper README - Hermes requires
context_length: 65536in config (tool schema overhead) - Thermal variance ±15% — use 3-pass eval before trusting keeps
Install
git clone https://github.com/jchacker5/autodecode.git
cd autodecode && uv sync