# NexaMass Runtime

`nexamass_encoder.py` is a minimal PyTorch loader for the model-state files in this repository.

Recommended checkpoint for inference:

```python
from runtime.nexamass_encoder import ModelConfig, load_nexamass_model_state

model = load_nexamass_model_state("weights/NexaMass-V3-Struct-model_state.pt", ModelConfig())  # PyTorch fallback
# Safetensors weights are also provided for safer tensor-only loading: weights/NexaMass-V3-Struct-model_state.safetensors
```

Use `weights/NexaMass-V3-Struct-model_state.safetensors` for the RDKit/Morgan-aligned model and `weights/Final_V3-model_state.safetensors` for the self-supervised foundation parent. PyTorch model-state `.pt` fallbacks are also included. Optimizer-bearing full training checkpoints remain in object storage and are intentionally excluded from this public HF payload.