[Feature] Portable Model Format

[mzau]

Summary

A self-contained model format optimized for distribution and consumption outside the HF ecosystem.

Motivation

• iOS/iPadOS: Import models via USB stick, AirDrop, or Files app - no HF SDK required
• macOS Finder: Quick Look preview, contextual menu integration ("Model Info")
• Offline sharing: Hand a USB stick to someone, they can use it immediately

Proposed Workflow

# 1. Clone to workspace (existing)
mlxk clone Qwen/Qwen3-8B ./models/qwen3

# 2. Export as portable model (new)
mlxk export ./models/qwen3 --portable ./portable/Qwen3-8B.mlxmodel

Portable Model Properties

• Single directory (or .mlxmodel bundle?)
• No symlinks, no .refs/, no blob hashes
• Minimal metadata: model.json with type, parameters, capabilities
• Weights as-is (safetensors)
• Optional: embedded icon, Quick Look generator

Open Questions

• Bundle format (.mlxmodel directory vs flat folder)?
• Finder integration scope (Quick Look only? Services menu?)
• iOS consumption API (Swift package? Documentation only?)

Non-Goals (for now)

• No target release
• Not a replacement for HF cache or workspace models
• No streaming/partial download support

Relation to Other Features

• Builds on upcoming workspace-local model features (2.0.5+)
• Complements mlxk clone / mlxk pull
• Could inform future mlxk convert enhancements

[mzau]

Status: Future / Community-driven

After discussion: The iOS angle requires Swift implementations of mlx-lm/mlx-vlm/mlx-audio - not just portable model files. That's a major undertaking beyond mlx-knife's Python scope.

This issue remains open as a vision for:

• macOS Finder integration (Quick Actions, Quick Look)
• Portable format for Mac-to-Mac sharing
• Foundation for a potential Swift port (if someone picks this up)

No target release. Contributions welcome if someone wants to explore this direction.