MADLAD-400 3B MT β€” GGUF (ggml)

GGUF / ggml conversion of google/madlad400-3b-mt for use with CrispStrobe/CrispASR.

MADLAD-400 is a 3B-parameter T5-based multilingual machine translation model covering 450+ languages, trained on 1 trillion tokens. It was developed by Google and achieves strong results across a wide range of languages, with special emphasis on low-resource and underrepresented languages. Distributed under Apache 2.0 license.

Files

File Size Notes
madlad400-3b-mt-f16.gguf ~5.7 GB F16 weights (reference quality)
madlad400-3b-mt-q8_0.gguf ~3.1 GB Q8_0 quantized
madlad400-3b-mt-q4_k.gguf ~1.8 GB Q4_K quantized

Quick start

# 1. Build CrispASR
git clone https://github.com/CrispStrobe/CrispASR
cd CrispASR
cmake -B build -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF
cmake --build build -j

# 2. Pull model
huggingface-cli download cstr/madlad400-3b-mt-GGUF madlad400-3b-mt-q8_0.gguf --local-dir .

# 3. Translate (uses <2xx> language tags)
./build/bin/crispasr --backend madlad -m madlad400-3b-mt-q8_0.gguf \
    --text "Hello world, how are you today?" \
    -sl en -tl de

# English β†’ Japanese
./build/bin/crispasr --backend madlad -m madlad400-3b-mt-q8_0.gguf \
    --text "Machine learning is changing the world." \
    -sl en -tl ja

# French β†’ Portuguese
./build/bin/crispasr --backend madlad -m madlad400-3b-mt-q8_0.gguf \
    --text "Bonjour le monde!" \
    -sl fr -tl pt

Supported languages (450+)

MADLAD-400 supports over 450 languages using <2xx> target language tags (ISO 639 codes). This includes all major world languages plus many low-resource languages. See the original model card for the full language list.

Architecture

Text β†’ SentencePiece tokenizer (256K vocab, shared encoder-decoder)
     β†’ <2xx> target language tag prepended to source text
     β†’ T5 encoder (24 layers, d=1024)
     β†’ T5 decoder (24 layers, d=1024) with cross-attention
     β†’ Greedy decode β†’ translated text

Conversion

python models/convert-madlad-to-gguf.py \
    --input google/madlad400-3b-mt \
    --output madlad400-3b-mt-f16.gguf

Related models

Downloads last month
4,069
GGUF
Model size
3B params
Architecture
t5
Hardware compatibility
Log In to add your hardware

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for cstr/madlad400-3b-mt-GGUF

Quantized
(25)
this model