Andromeda (2026)

OMNI Andromeda (2026)

The first public generation of OMNI, a lossless source-code compression system designed specifically for structured programming code.

Andromeda combines learned neural prediction with traditional compression techniques, allowing OMNI to choose between neural literal encoding and structural LZ matches based on estimated coding cost.

Highlights

• Lossless, byte-exact reconstruction
• Reversible Python source tokenization
• BPE for identifiers, strings, and comments
• Continuous GRU-based source-code prediction
• Cost-based neural-vs-LZ representation selection
• Entropy-coded LZ metadata
• Arithmetic coding
• Designed for Python source repositories

Benchmark

On 9 completely unseen Python repositories, Andromeda produced smaller compressed outputs than tar+xz on all 9:

• 9/9 repositories beat tar+xz
• Average advantage: ~2.8 percentage points
• Best advantage: +6.2 percentage points
• 360/360 files reconstructed byte-for-byte

The 21.2 MB model is a shared model artifact and is intended to be amortized across multiple repositories/files rather than charged against an individual archive.

Included artifacts

• neural_model_andromeda.pt Andromeda neural model
• libomni.so native OMNI runtime

Andromeda is the 2026 generation of the SATISH format family and is associated with the .satish_andromeda extension.

Future OMNI generations will build upon previous models and expanded training data while maintaining compatibility with previously released SATISH archives.