maderix

maderix / ANE

Public

Training neural networks on Apple Neural Engine via reverse-engineered private APIs

115
15
80% credibility
Found Mar 01, 2026 at 40 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Objective-C
AI Summary

This repository provides tools to train small transformer language models, like Stories110M, directly on Apple's Neural Engine hardware in Apple Silicon Macs using custom low-level programs.

How It Works

1
🔍 Discover ANE Training

You stumble upon this exciting project on GitHub that lets you train AI to write stories using the hidden power in your Apple Mac's chip.

2
💻 Prepare your Mac

Make sure you have a recent Apple Silicon Mac, download the files, and get some story data ready with a quick script.

3
🔨 Build the trainer

Use a single command in your terminal to create the training program from the files.

4
🚀 Launch the training

Start the program and feel the thrill as it begins teaching the AI to understand and create simple stories right on your Mac's special hardware.

5
📊 Watch it learn

Fire up the colorful dashboard to track progress, see the learning curve improve, and monitor your Mac's energy use.

🎉 Enjoy story generation

Celebrate as your AI model gets smarter over time and starts generating fun, tiny stories from what it learned.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 40 to 115 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is ANE?

ANE lets you train transformer models directly on Apple's Neural Engine using reverse-engineered private APIs in Objective-C. It handles full forward and backward passes on ANE hardware—no CoreML training, Metal, or GPU needed—delivering 9.3 ms per step for a 768-dim layer on M4 at 1.78 TFLOPS. Build with a single clang command on macOS 15+ Apple Silicon, run binaries like train_large on TinyStories data, and monitor via a Python dashboard showing loss curves, power draw, and generated text.

Why is it gaining traction?

It cracks open ANE for training, a 15.8 TFLOPS accelerator Apple locks to inference, with optimizations hitting 11% utilization on multi-layer models like 110M-param Stories110M at 107 ms/step. The dashboard tracks ANE power, CPU/memory, and auto-generates samples, making experiments like github repo training or training neural machine translation with terminology constraints feel like a polished github training app. Low compile overhead via exec restarts keeps it snappy.

Who should use this?

Apple Silicon ML engineers prototyping on-device training for small LLMs. Researchers benchmarking ANE vs GPU, like anemia in hardware utilization or aneurysma in compute graphs. Devs running github training data locally without cloud, especially for seq=256 stories or custom vocab setups.

Verdict

Grab it for Apple-and-Github experiments if you're on M4+—dashboard and checkpoints make it fun despite 18 stars signaling early stage. 0.800000011920929% credibility fits the private-API risks; solid docs but test your own gradients before prime time.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.