FIND-20260323-013 — maderix/ANE: Training on Apple Neural Engine via reverse-engineered private APIs

Summary

maderix/ANE — Training neural networks on Apple Neural Engine via reverse-engineered private APIs

Developer Manjeet Singh reverse-engineered Apple's undocumented _ANEClient and _ANECompiler private APIs to enable full transformer training — forward pass, backward pass, attention, gradients, and optimizer — running directly on Apple's Neural Engine. No CoreML. No Metal. No GPU. Pure ANE compute.

Achieves 9.3ms per training step on M4 silicon and 6.6 TFLOPS/W power efficiency. Builds with a single clang command and zero external dependencies (macOS system frameworks only).

6,444

GitHub Stars

9.3ms

Per training step (M4)

15.8

TFLOPS raw (ANE)

External dependencies

Research caveat: The ANE compiler leaks resources and crashes after approximately 119 compilations. The workaround is self-respawning via exec(). Not production-ready without significant hardening.

apple neural-engine on-device-ai reverse-engineering objective-c machine-learning macos tauri desktop-ai trending

Security Review

License

MIT

Last Commit

2026-03-10

Known CVEs

Maintenance

MAINTAINED

Supply Chain Risk

LOW

Recommendation

USE_WITH_CAUTION

Zero external dependencies — macOS system frameworks only. Single maintainer (maderix). No published security policy. USE_WITH_CAUTION due to explicit instability constraints (resource leak, ~119 compilation limit, exec() self-restart workaround).

ODS Impact

ODS runs on Linux/GCP — ANE hardware is not directly applicable to server-side services. However, two strategic angles merit tracking:

DocSign (Tauri/macOS): DocSign is a Tauri desktop app targeting macOS. Future on-device AI capabilities — document intelligence, signature verification, content analysis — could leverage ANE training for local fine-tuning without cloud round-trips. 15.8 TFLOPS with no network dependency is compelling for a privacy-first document product.
Zero-dependency ML build pattern: The single-clang-command, zero-package-manager build model is worth noting for embedding lightweight AI into Rust/native desktop clients. Contrasts sharply with PyTorch/TensorFlow dependency hell.

Not actionable for ODS server-side services today. Monitor for maturation into a production-usable framework before considering DocSign integration.