MLX Model Manager

MLX Model Manager provides a unified interface for loading and running both Large Language Models (LLMs) and Vision-Language Models (VLMs) on Apple Silicon. It simplifies the model download, initialization, and generation process, so you can focus on integrating AI capabilities into your applications.

Paligemma2-4k.mov

Installation

📦 Use Swift Package Manager to integrate MLX Model Manager into your project.

1. Open Xcode and go to File > Add Packages....
2. Enter the repository URL for mlx-model-manager.
3. Add MLXModelManager to your target.
4. Use import MLXModelManager in your Swift project.

Usage

func loadAndGenerate() async {
    let manager = ModelManager(modelPath: "mlx-community/phi-2-hf-4bit-mlx")
    do {
        // Load the model
        try await manager.loadModel()
        print("Model loaded successfully.")
        
        // Generate text
        await manager.generate(prompt: "What is the capital of France?")
        print("Generated Output: \(manager.output)")
    } catch {
        print("Error: \(error)")
    }
}

Example: Loading an LLM

import SwiftUI
import MLXModelManager

struct ContentView: View {
    @StateObject var manager = ModelManager(modelPath:"mlx-community/phi-2-hf-4bit-mlx")

    var body: some View {
        VStack(spacing: 20) {
            Text("🤖 LLM Example").font(.headline)
            Button("Load & Generate") {
                Task {
                    try await manager.loadModel()
                    await manager.generate(prompt: "What is the capital of France?")
                }
            }
            ScrollView {
                Text(manager.output).padding()
            }
        }
        .padding()
    }
}

Example: Loading a VLM

import SwiftUI
import MLXModelManager

struct ContentView: View {
    @StateObject var manager = ModelManager(modelPath: "mlx-community/paligemma2-3b-ft-docci-448-8bit")

    var body: some View {
        VStack(spacing: 20) {
            Text("🖼️ VLM Example").font(.headline)
            Button("Load & Describe Image") {
                Task {
                    try await manager.loadModel()
                    await manager.generate(
                        prompt: "Describe the image in English.",
                        imagePath: "/path/to/your/image.png"
                    )
                }
            }
            ScrollView {
                Text(manager.output).padding()
            }
        }
        .padding()
    }
}

Parameters

⚙️ You can adjust various parameters before generating text:

• temperature: Controls randomness. Lower values = more deterministic output.
• topP: Nucleus sampling threshold. Consider only tokens within cumulative probability topP.
• repetitionPenalty: Penalizes repeated tokens to encourage diversity.
• maxTokens: Limits the maximum number of tokens to generate.

Set parameters before calling generate:

manager.maxTokens = 200
manager.setHyperparameters(temperature: 0.7, topP: 0.9, repetitionPenalty: 1.0)
try await manager.loadModel()
await manager.generate(prompt: "Explain quantum mechanics simply.")

Supported Models

You can replace the modelPath parameter in your Swift code with any of the IDs listed below. All models support Language modality, and the last two also support Vision.

ModelPath	Language	Vision
mlx-community/SmolLM-135M-Instruct-4bit	✅
mlx-community/Mistral-Nemo-Instruct-2407-4bit	✅
mlx-community/Mistral-7B-Instruct-v0.3-4bit	✅
mlx-community/CodeLlama-13b-Instruct-hf-4bit-MLX	✅
mlx-community/phi-2-hf-4bit-mlx	✅
mlx-community/Phi-3.5-mini-instruct-4bit	✅
mlx-community/Phi-3.5-MoE-instruct-4bit	✅
mlx-community/phi-4-4bit	✅
mlx-community/quantized-gemma-2b-it	✅
mlx-community/gemma-2-9b-it-4bit	✅
mlx-community/gemma-2-2b-it-4bit	✅
mlx-community/Qwen1.5-0.5B-Chat-4bit	✅
mlx-community/OpenELM-270M-Instruct	✅
mlx-community/Meta-Llama-3.1-8B-Instruct-4bit	✅
mlx-community/Meta-Llama-3-8B-Instruct-4bit	✅
mlx-community/Llama-3.2-1B-Instruct-4bit	✅
mlx-community/Llama-3.2-3B-Instruct-4bit	✅
mlx-community/paligemma-3b-mix-448-8bit	✅	✅
mlx-community/Qwen2-VL-2B-Instruct-4bit	✅	✅
mlx-community/paligemma2-3b-ft-docci-448-8bit	✅	✅

Attribution

This is built on top of mlx-swift-examples. Originally developed by davidkoski, awni, and the MLX team, with additional contributions from the community. Special thanks to Prince Canuma, who created MLX-VLM, later ported to Swift by the MLX team.

MLX Model Manager unifies how you can interact with both LLMs and VLMs, streamlining development and integration efforts.