Zyquo MLX
The on-device model foundry for Apple Silicon
Run, fine-tune, quantize, and convert LLMs 100% locally with MLX — the complete model workbench, nothing leaves your machine.
4
fine-tuning methods (LoRA, QLoRA, DoRA, full)
604–630
tok/s inference (Qwen3-0.6B, M5 Max)
0.1%
quantization size-preview accuracy
~3,300
tok/s LoRA training throughput
1.0 s
to transcribe a 7 s clip (Whisper large-v3-turbo)
100%
local — zero telemetry
Overview
Zyquo MLX is the foundry of the Zyquo family: where Zyquo Local is the chat client, MLX is the complete workbench for the on-device model lifecycle on Apple Silicon. It runs every MLX model type — streaming LLM chat, vision-language models with images, embeddings with a live similarity inspector, and Whisper speech-to-text — with tokens/sec, time-to-first-token, and verified memory release on every run.
Fine-tuning is first-class: LoRA, QLoRA, DoRA, and full fine-tuning through a real configurator with live loss curves, checkpoints, cancel and warm resume, and memory gating that blocks impossible configs before they run. Quantization to 4 or 8 bits ships size previews accurate to 0.1%, adapter fusing with a smart de-quantize default, and Hugging Face to MLX conversion.
Everything stays local — no API keys, no telemetry; the only network traffic is downloading the models you ask for. A live-verified Featured catalog and full mlx-community search come with RAM compatibility badges for your specific Mac, and datasets get row-by-row JSONL validation with concrete fixes and deterministic train/valid splits.
Key Features
Every MLX model type
Streaming LLM chat, vision-language models with images, embeddings with a live similarity inspector, and Whisper speech-to-text — all on-device.
Fine-tune on your data
LoRA, QLoRA, DoRA, and full fine-tuning with live loss curves, checkpoints, cancel and warm resume, and a real training configurator.
Memory-gated training
Impossible training configurations are blocked before they run, based on your Mac's actual memory — no mid-run out-of-memory surprises.
Quantize and convert
4/8-bit affine quantization with size previews accurate to 0.1%, adapter fusing with a smart de-quantize default, and Hugging Face to MLX conversion.
Datasets done right
Import JSONL in chat, prompt-completion, or text formats with row-by-row validation, concrete fixes, deterministic train/valid splits, and token stats.
Model discovery with RAM badges
A live-verified Featured catalog plus full mlx-community search, stamped with RAM compatibility badges for your Mac, and resumable downloads.
Side-by-side evaluation
Compare the base model against your fine-tune on the same prompt, with per-side generation statistics.
Measured, verified performance
Tokens/sec, time-to-first-token, and verified memory release on every run; quantization predicted 335.3 MB and produced 335.5 MB.
Nothing leaves your machine
No API keys and no telemetry — the only network traffic is downloading the models you explicitly request.
How It Works
InferenceEngine actor
A Swift actor handling LLM, VLM, and embedding inference natively via mlx-swift, with a MemoryAdvisor gating loads and a SpeechService for Whisper transcription.
Training layer
TrainingService, RunStore, and MetricsStream drive fine-tuning runs through a live JSON protocol — loss curves, checkpoints, and warm resume all flow through it.
PyBridge
A PythonRunner drives a pinned mlx-lm 0.31.3 in an isolated, uv-provisioned venv through strict JSON-lines scripts — Python is an implementation detail, never the interface.
Convert and Data services
ConversionService combines Swift-native quantization with Python fuse/convert paths; DatasetService validates JSONL row-by-row, splits deterministically, and previews token stats.
Hub layer
HubService, a resumable DownloadManager, and ModelStore handle catalog discovery, mlx-community search, and verified model delivery.
Tech Stack
App
Inference
Training
Quality
Highlights
- Quantization size predictions accurate to 0.1% — a 1.19 GB fp16 model predicted at 335.3 MB came out at 335.5 MB
- QLoRA training at ~2,300 tok/s with a 0.8 GB peak; LoRA at ~3,300 tok/s — measured on an M5 Max
- Works around two documented upstream mlx-lm landmines, with the research published in docs/TRAINING-RESEARCH.md
- Swift-native inference and quantization; Python only for training and speech, pinned and sandboxed behind a JSON protocol
- Built without the Xcode IDE — command-line only, no .xcodeproj
- Signed and notarized with a Developer ID; MIT licensed