Vāķ Translate

Vāķ is Shunya's open-weight translation model for Indian languages, 55 languages, 2,970 any-to-any pairs, and a weighted average BLEU of 38.5, with native handling of all major Indic scripts. This page documents vak-translate-1.3b-ct2, the CTranslate2 build of vak-translate-1.3b for fast local inference, released under CC-BY-SA-4.0 as part of the Vāķ suite from Shunya Labs at the India AI Impact Summit 2026.

Highlights

• 55 Indian languages across 5 language families (Indo-Aryan, Dravidian, Austroasiatic, Sino-Tibetan, Indo-European)
• 2,970 translation pairs, any-to-any translation between all supported languages
• 1.3B parameters, encoder-decoder architecture with 24+24 layers
• Open weights under CC-BY-SA-4.0
• Weighted average BLEU: 38.5 (by speaker count)
• Covers 1.17 billion+ native speakers across every region of India
• First open-weight translation model for many Indian languages including Bhojpuri, Rajasthani, Chhattisgarhi, and Magahi
• CTranslate2 format for optimized CPU/GPU inference

At a glance

Model architecture

Supported languages

Full language list with BLEU scores (from the model card). BLEU scores are tentative, based on human evaluation (3 independent evaluations per language, 1-5 adequacy scale).

Language families

Performance tiers

Evaluation

• Weighted Average BLEU (by speaker count): 38.5
• BLEU scores are tentative, based on human evaluation (3 independent evaluations per language, 1-5 adequacy scale)
• Covers 1.17 billion+ native speakers across 5 language families

Translate on your machine

Install a few Python packages, run the script from the Hugging Face quick start, and read the translated sentence in your terminal. No cloud API required.

What goes in and what comes out

Same example as the model card: one English sentence in, one Hindi sentence out.

The Hindi line is what print(translation) shows when you run the script below with this English input. Wording may vary slightly between runs.

Install

pip install ctranslate2 transformers sentencepiece

CTranslate2 quick start

import ctranslate2
from transformers import NllbTokenizer
from huggingface_hub import snapshot_download

# Download model to local cache and load
model_dir = snapshot_download("shunyalabs/vak-translate-1.3b-ct2")

tokenizer = NllbTokenizer.from_pretrained("shunyalabs/vak-translate-1.3b-ct2")

device = "cuda" if ctranslate2.get_cuda_device_count() > 0 else "cpu"
translator = ctranslate2.Translator(model_dir, device=device)

# Translate English to Hindi
src_lang = "eng_Latn"
tgt_lang = "hin_Deva"

tokenizer.src_lang = src_lang
inputs = tokenizer("Hello, how are you?")
src_tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"])

results = translator.translate_batch(
    [src_tokens],
    target_prefix=[[tgt_lang]],
    beam_size=4,
    max_decoding_length=256,
)

output_tokens = results[0].hypotheses[0]
output_ids = tokenizer.convert_tokens_to_ids(output_tokens)
translation = tokenizer.decode(output_ids, skip_special_tokens=True)
print(translation)

Translate many sentences at once

From the model card, pass a list of English sentences; the script prints one Hindi line per sentence.

# Translate a batch of sentences (English → Hindi)
texts = [
    "The sun rises in the east.",
    "Water is essential for life.",
    "Education is the most powerful weapon.",
]

tokenizer.src_lang = "eng_Latn"
all_src_tokens = [
    tokenizer.convert_ids_to_tokens(tokenizer(t)["input_ids"])
    for t in texts
]

results = translator.translate_batch(
    all_src_tokens,
    target_prefix=[["hin_Deva"]] * len(texts),
    beam_size=4,
    max_decoding_length=256,
)

for orig, result in zip(texts, results):
    ids = tokenizer.convert_tokens_to_ids(result.hypotheses[0])
    print(tokenizer.decode(ids, skip_special_tokens=True))

Language codes

Vāķ uses NLLB-style language codes: <iso-639-3>_<iso-15924>. Examples from the quick start:

Use cases

Combining with ASR + TTS

Vāķ is the translation bridge for live 2-way conversation. The dashboard's live-translation feature (and any voice agent you build) uses it in a pipeline like this: