v0.1.0 | Apache 2.0 | Powered by HammerIO

Your models. Your hardware. Zero cloud.

Compress, version, store, verify, and restore AI model checkpoints locally. A 7B model checkpoint restores in 3.6 seconds on Jetson AGX Orin.

Live Dashboard

ModelSafe dashboard showing model checkpoint management

ModelSafe — Model checkpoint management dashboard

Benchmark Results

3.6s 7B model restore time
7.9s 13B model restore time
391 MB/s Peak decompress throughput
100% SHA-256 integrity — all tests PASS

Checkpoint Benchmark Results

Model Size Original Compressed Ratio Compress Decompress Restore Time Integrity
1GB (7B equiv) 1.0 GB 875.1 MB 1.17x 208 MB/s 283 MB/s 3.615s PASS
3GB (13B equiv) 3.0 GB 2.4 GB 1.24x 237 MB/s 391 MB/s 7.861s PASS
7GB (70B equiv) 7.0 GB 5.7 GB 1.23x 289 MB/s 378 MB/s 18.979s PASS

Synthetic model weight data — float32 arrays with realistic distribution. Ratios reflect actual model weight compressibility. Restore time includes SHA-256 verification. Powered by HammerIO nvCOMP GPU LZ4.

How It Works

ModelSafe four-stage pipeline — Store with SHA-256 hash, Compress via nvCOMP GPU LZ4, Verify integrity on every restore, Restore in 3.6 seconds

Use Cases

  • Field node recovery

    Restore inference capability after reboot in seconds.

  • Model versioning

    Tag and track checkpoint versions locally.

  • Bandwidth-constrained deployment

    Compressed transfer over limited links.

  • Air-gapped archival

    Secure local checkpoint storage with integrity guarantees.

Integration

ModelSafe uses HammerIO as its compression engine and reports restore telemetry to PraetorianMind. AriaOS: Forge model checkpoints are automatically compatible with ModelSafe storage.

Explore the full platform

Explore the full platform →