MLX-quantized open-weight LLMs, CoreML detectors, and the exact models that ship with claude-code-local — Matt Macosko's on-device AI hub.
Muse-Glimmer 30B Abliterated went up here on August 12. Two days later it is already the fifth most-downloaded model on this page, out of fifteen. A multimodal 30B that runs on an M-series Mac, because the mlx-vlm architecture support went in the same week the model dropped.
The Nice Dreamz software portfolio is mostly the code — agents, browsers, recorders, dashboards. This page is the weights: the actual LLMs and vision models I quantize, convert, and ship for Apple Silicon. They power claude-code-local (3,249 ⭐ on GitHub), the on-device RealTime AI Camera iOS app, and the AirGap AI consulting pilot for legal/healthcare workflows.
Everything below is open-weight. The lineup pulls about 6,500 downloads every 30 days — almost entirely by people running them on their own M-series Macs through mlx-lm. No cloud inference, no API keys, no telemetry.
The pitch for AirGap AI — the consulting practice — is simple: privileged documents in, answers out, never a byte to a third-party cloud. That only works if the weights live on the firm's hardware. Cloud models are off the table.
So I quantize and convert the best open models for Apple Silicon, publish them here, and ship them inside claude-code-local. The same files a law firm uses in their air-gapped pilot are the same ones any developer can huggingface-cli download right now.
The fastest path: install claude-code-local — it pulls the right weights from this Hugging Face profile automatically. For confidential workflows (legal, healthcare), the AirGap AI pilot ships the whole stack into a real air-gapped environment with verified network audits.