ncdrone

ncdrone / rustane

Public

Rust-native hybrid inference engine for Apple Neural Engine + Metal GPU

20
2
100% credibility
Found Mar 20, 2026 at 20 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Rust
AI Summary

Rustane enables everyday Mac users to train and test large AI language models directly on their Apple hardware for efficient, low-power experimentation.

How It Works

1
🔍 Discover rustane

You find this project on a coding site and get excited about training smart AI models right on your Apple computer without needing fancy servers.

2
💻 Check your setup

See if your Mac has enough memory (like 18GB or more) to handle the fun sizes of AI models it supports.

3
⚡ Run your first test

With one simple command, build and test a tiny AI trainer to confirm everything works perfectly on your machine.

4
📊 Benchmark speeds

Try quick runs on models from small to huge to measure how fast your Mac trains them and find your sweet spot.

5
🚀 Train a real model

Pick a model size that fits your memory, feed it some text data, and watch it learn step by step.

🎉 Enjoy fast results

Your AI model trains super efficiently using your Mac's special chips, and you share your speed scores with the community.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 20 to 20 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 rustane?

Rustane is a Rust-native hybrid inference and training engine for Apple's Neural Engine and Metal GPU, letting you train transformer models up to 5B parameters or run forward passes to 30B on M4 Max hardware. It compiles custom kernels directly to ANE via private APIs, skipping CoreML black boxes, and exports trained weights as SafeTensors for anywhere inference. Run `make sweep-600m` to validate a full training pipeline in 17 seconds on any 18GB+ Apple Silicon Mac.

Why is it gaining traction?

It trains at 3-5W power draw, freeing the GPU entirely, with architecture sweeps finding optimal depth/width configs across scales in under an hour. Benchmarks scale predictably by RAM—no ANE compute ceiling hit—and community results pour in via GitHub issues. Rust-native safety plus reverse-engineered ANE access beats Python wrappers for low-level control on Apple hardware.

Who should use this?

ML engineers on Apple Silicon training small-to-mid LLMs locally, like 600M-5B param models on climbmix datasets. Researchers sweeping transformer variants (wide/shallow vs deep/narrow) without cloud costs. Rust devs prototyping hybrid ANE+Metal pipelines for on-device inference.

Verdict

Grab it if you're on M-series chips chasing efficient local training—benchmarks are plug-and-play solid. At 20 stars and 1.0% credibility, it's early and reverse-engineered (private APIs may shift), so test thoroughly before production.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.