The first release of Zyquo MLX — the local MLX foundry for the Mac. Run, fine-tune, quantize, convert, and evaluate models entirely on Apple Silicon. Nothing leaves your machine.
✨ Highlights
- Inference for every MLX model type: streaming LLM chat, vision-language models with images, embeddings with a similarity inspector, Whisper speech-to-text — with live tok/s, TTFT, and verified memory release
- Real fine-tuning: LoRA / QLoRA / DoRA / full, with live loss curves, checkpoints, cancel + warm resume, and memory gating calibrated to your Mac
- Quantize & convert: 4/8-bit affine quantization (size previews accurate to 0.1 %), adapter fusing with an adapter-preserving de-quantize default, Hugging Face → MLX conversion
- Datasets: JSONL import with row-level validation and fixes, deterministic splits, token stats
- Discover: live-verified Featured catalog + full mlx-community search with RAM badges and resumable downloads
- Evaluate: base vs. fine-tuned side by side
📦 Installation
- Download
ZyquoMLX.dmgbelow (signed & notarized — Developer ID) - Drag Zyquo MLX to Applications and launch
- Grab a model from Models › Discover
✅ Requirements
- Apple Silicon Mac (M1 or later), macOS 14+
- For fine-tuning/conversion/speech:
brew install uv(isolated Python env is provisioned automatically on first use); speech also usesbrew install ffmpeg
Measured performance and the full verification matrix: docs/VERIFICATION.md