initial release: deploy code + split scripts

docs: initial README

Add Gemma 4-26B-A4B support: 4.15 tok/s on M4 Mac Mini

Add Gemma 4 MLX model class + preprocess

Add Gemma 4 MLX model class + preprocess

Add Gemma 4 MLX model class + preprocess

Add Gemma 4 MLX model class + preprocess

Add Gemma 4 MLX model class + preprocess

Add Coder model + multi-model fixes

Add Coder model + multi-model fixes

Add Coder model + multi-model fixes

Add Coder model + multi-model fixes

Add Coder model + multi-model fixes

Add chat command + shared generate module

Add chat command + shared generate module

Add routing bias comparison table: bias=1.0 universal sweet spot

Update 30B: 4.29 tok/s with bias=1.0

Add 30B support + bias sweep results

Add 30B support + bias sweep results

Add 30B support + bias sweep results

Add Ollama-compatible serve command

Add Ollama-compatible serve command

Add download command: mlx-sniper download qwen3.5-35b

Add download command: mlx-sniper download qwen3.5-35b

Add supported models table and hardware requirements

Update src/mlx_expert_sniper/preprocess.py

Update src/mlx_expert_sniper/expert_io.py

Update src/mlx_expert_sniper/engine.py

Update src/mlx_expert_sniper/coactivation.py

Update src/mlx_expert_sniper/calibrate.py

Update src/mlx_expert_sniper/cli.py

Update src/mlx_expert_sniper/__init__.py

Update setup.py

Update setup.cfg

Update pyproject.toml

Correct to 5.37 tok/s, remove REAP from claims

35B: 5.18 tok/s with REAP masking + routing bias=1.0

Update verified results: 35B at 2.42 tok/s, 30B at 3.34 tok/s

Update benchmark numbers: verified 3.3 tok/s across varied prompts

Update README.md

Add YAML metadata to fix repo card warning

v0.3.0: Add Qwen3-Coder-30B + qwen3_5_moe support, thinking filter, 3 models verified

v0.2.0: Add Qwen3.5-35B-A3B support (5.78 tok/s, 19.5 GB on 16 GB RAM)

v0.1.0: MoE expert sniping for MLX — run models larger than your RAM

initial commit