mlx-model-lab

MLX model conversion, quantization, and benchmarking toolkit for Apple Silicon.

Converted models are published to HuggingFace mlx-community.

What this repo contains

mlx-model-lab/
├── scripts/          # Automation: convert, benchmark, upload, monitor
├── benchmarks/       # Results by model (JSON + markdown reports)
├── model-cards/      # Generated model card READMEs for HuggingFace
├── configs/          # Quantization configs (OptiQ, oQ profiles)
└── docs/             # Guides, methodology notes

Model weights are not stored here — they live on HuggingFace.

Hardware

• Apple M4 Pro, 48GB unified memory, 1TB SSD
• macOS 26 (Tahoe)
• oMLX v0.3.5

Workflow

1. Identify gap       → scripts/scan-gaps.py (check HF for missing MLX models)
2. Convert            → scripts/convert.sh <model> <bits>
3. Validate quality   → scripts/benchmark.sh <model> (perplexity + evals)
4. Generate card      → scripts/gen-model-card.py <model> (from benchmark data)
5. Upload             → scripts/upload.sh <model> mlx-community/<name>

Quantization methods

Method	When to use
Standard (mlx_lm.convert -q)	Quick first conversion, always do 4bit + 8bit
AWQ (mlx_lm.awq)	Better quality, moderate speed
DWQ (mlx_lm.dwq)	Best weight quality for the bit budget
GPTQ (mlx_lm.gptq)	Per-layer optimization
OptiQ (mlx-optiq)	Mixed-precision: 4bit for robust layers, 8bit for sensitive
oQ (oMLX built-in)	Data-driven mixed-precision, best at very low bits (2-3bit)

License

MIT