mlx-community — Community MLX-Swift Implementations

Hey everyone! 👋

I've been building MLX-Swift implementations of recent research papers and wanted to share them with the community. I've organized these under a new GitHub org: github.com/mlx-community

The goal is to be a community-driven complement to ml-explore — focusing on bleeding-edge techniques and native Swift implementations that aren't in the official repos yet.

---

Current Projects

🚀 speculative-decoding

Native MLX-Swift implementation of speculative decoding for faster LLM inference.

• Draft-then-verify pipeline with any MLX-compatible models
• 2-3x speedups on Apple Silicon
• Streaming support, simple API
• Works with models from mlx-community on HF

let output = try await SpeculativeDecoding.generate(
    prompt: "Explain quantum computing:",
    draftModelId: "mlx-community/Qwen2.5-0.5B-Instruct-4bit",
    targetModelId: "mlx-community/Qwen2.5-7B-Instruct-4bit"
)

Benchmarks (M2 Max, 24GB):

Method	Tokens/sec	Speedup
Standard	45.2	1.0x
Speculative (warm)	89.4	1.98x

---

🧠 nano-reasoning

FastRL + EAGLE adaptive drafter implementation. The drafter learns from target model rejections in real-time, improving acceptance rate during inference.

Based on:

• EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty
• FastRL: Adaptive Drafter for Efficient Reasoning

Features:

• Hardware-tiered configs (Entry/Pro/Elite for different Mac specs)
• Background training with GPU-aware scheduling
• Single-layer EAGLE head option for minimal overhead
• Multi-Armed Bandit for adaptive SD configuration

---

📄 paddleocr-vl-swift

Native Swift port of PaddleOCR-VL — a 0.9B vision-language model for document understanding.

• OCR, table parsing, formula extraction, chart analysis
• 109 language support
• NaViT-style dynamic resolution
• CLI + library interfaces

paddleocr-vl ocr document.png --task table

---

Roadmap

Working on next:

• mamba-swift — Native Mamba/Mamba-2 SSM (no Swift implementation exists yet)
• medusa-swift — Multi-head speculative decoding
• lora-swift — LoRA training in native Swift

---

Why MLX-Swift?

Most MLX community work is Python. But there's a huge opportunity for native Swift:

• Ship ML directly in iOS/macOS apps
• No Python runtime overhead
• Better integration with Apple ecosystem
• SwiftUI + MLX = native AI apps

---

Happy to take suggestions on what to implement next. PRs welcome!

— @lulzx