---
license: mit
tags:
- colbert
- onnx
- zero-pytorch
- modernbert
- text-embeddings
pipeline_tag: feature-extraction
library_name: generic
---

# lateon-onnx

This repository hosts the optimized, single-file **ONNX representation** of the ModernBERT-backed [lightonai/LateOn](https://huggingface.co/lightonai/LateOn) model.

It is designed to run completely PyTorch-free and dependency-free using the [intextus-embed](https://github.com/Intextus/intextus-embed) runtime library.

## Model Metadata
- **Backbone**: ModernBERT-base (140M parameters)
- **Output Dimensions**: 128-dimensional late-interaction embeddings
- **ONNX File Size**: 580 MB (fully self-contained, merged weights)
- **Case Sensitivity**: Case-sensitive (requires `do_lower_case=False`)

## Usage

Install the `intextus-embed` runtime:
```bash
pip install intextus-embed
```

Load the model automatically and run inference (set `do_lower_case=False` because LateOn is case-sensitive):
```python
from intextus import IntextusEncoder, compute_maxsim

# Automatically downloads and caches the model from Hugging Face
model = IntextusEncoder("lateon", do_lower_case=False)

# Encode queries and documents
query_embeddings = model.encode_queries("What is ultra-low latency?")
doc_embeddings = model.encode_docs("ONNX runtime bypasses the PyTorch layer completely.")

# Compute the MaxSim similarity score via NumPy
score = compute_maxsim(query_embeddings[0], doc_embeddings[0])
print(f"Relevance Score (MaxSim): {score:.4f}")
```