SuperInstance

Micro-model ecosystem: distill LLM knowledge into deployable micro-models.

Packages

| Package | PyPI | npm | crates.io | Description | |---------|------|-----|-----------|-------------| | plato-core | pip install plato-core | @superinstance/plato-core | — | Base types + mesh registry | | tensor-spline | pip install tensor-spline | @superinstance/tensor-spline | — | SplineLinear layers, 5-20x compression | | eisenstein-embed | pip install eisenstein-embed | @superinstance/eisenstein-embed | — | 5-layer matching cascade | | plato-training | pip install plato-training | — | — | Training framework (monolith) | | plato-deadband | — | — | plato-deadband | Deadband caching (Rust) | | constraint-theory-core | — | — | constraint-theory-core | Constraint solving (Rust) | | spectral-conservation | — | — | spectral-conservation | Spectral analysis (Rust) |

Architecture

Each package is standalone — install only what you need. When co-installed, packages auto-discover and mesh via entry_points.

See MESH-ARCHITECTURE.md for the full specification.

Key Results

• Eisenstein encoder: 71.2% hit rate, 653x smaller than Model2Vec
• SplineLinear: 16,384:1 compression ratio on 512×512 layers
• Bitvector matching: 93.8% typo accuracy, zero ML dependencies
• ONNX inference: 58,648 qps on CPU (700x faster than PyTorch)
• Heterogeneous compute: CUDA (training) + CPU/ONNX (inference) + iGPU (overflow) + NPU (pending)