---
license: apache-2.0
base_model: google/madlad400-3b-mt
tags:
  - translation
  - onnx
  - int8
  - multilingual
  - nobuds
language:
  - multilingual
---

# MADLAD-400 3B ONNX INT8

INT8 quantized ONNX export of [google/madlad400-3b-mt](https://huggingface.co/google/madlad400-3b-mt) for on-device multilingual translation.

- **Parameters:** 3B (T5-based)
- **Languages:** 450+
- **License:** Apache 2.0
- **Quantization:** INT8 dynamic (optimum O3)
- **Export tool:** optimum-cli

## Files

- `encoder_model_quantized.onnx` — Encoder
- `decoder_model_quantized.onnx` — Decoder (initial)
- `decoder_with_past_model_quantized.onnx` — Decoder (with past key-values)
- `spiece.model` — SentencePiece tokenizer (T5)
- `tokenizer.json` — Fast tokenizer config

## Language prefix

MADLAD-400 uses target-language prefixes: `<2xx>` where `xx` is the language code.
Example: `<2de> Hello world` translates to German.

## Note

At ~3 GB INT8, this model is large for mobile deployment. Consider M2M-100 418M (~630 MB) for resource-constrained devices.