shunyalabs/vak-translate-1.3b-ct2

Vāķ Translate 1.3B — CTranslate2

CTranslate2-converted version of vak-translate-1.3b — open-weight translation model covering 55 Indian languages and 2,970 language pairs.

This is the CTranslate2 optimized version of the original model, enabling faster CPU and GPU inference with reduced memory usage. Converted from the original Transformers checkpoint.

Built by Shunya Labs. Part of the Vāķ suite launched at the India AI Impact Summit 2026.

CTranslate2 Quick Start

pip install ctranslate2 transformers sentencepiece

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)

Batch Translation

# 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))

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

Supported Languages

Full Language List with BLEU Scores

#	Language	Speakers	BLEU	#	Language	Speakers	BLEU
1	Hindi	322.2M	42	28	Pahari	3.25M	20
2	Bengali	96.2M	41	29	Bhili	3.21M	23
3	Marathi	82.8M	40	30	Harauti	2.94M	23
4	Telugu	80.9M	41	31	Nepali	2.93M	36
5	Tamil	68.9M	42	32	Bagheli	2.68M	34
6	Gujarati	55.0M	40	33	Sambalpuri	2.63M	23
7	Urdu	50.7M	41	34	Dogri	2.60M	3
8	Bhojpuri	50.6M	36	35	Garhwali	2.48M	35
9	Kannada	43.5M	40	36	Nimadi	2.31M	26
10	Malayalam	34.8M	41	37	Konkani	2.15M	15
11	Odia	34.1M	39	38	Kumauni	2.08M	34
12	Punjabi	31.1M	40	39	Kurukh	1.98M	3
13	Rajasthani	25.8M	36	40	Tulu	1.84M	3
14	Chhattisgarhi	16.3M	32	41	Manipuri (Meitei)	1.76M	3
15	Assamese	14.8M	38	42	Surgujia	1.74M	28
16	Maithili	13.4M	37	43	Sindhi	1.68M	35
17	Magahi	12.7M	35	44	Bagri	1.66M	12
18	Haryanvi	9.81M	23	45	Ahirani	1.64M	34
19	Khortha	8.04M	34	46	Banjari	1.58M	34
20	Marwari	7.83M	36	47	Brajbhasha	1.56M	35
21	Santali	6.97M	3	48	Bodo	1.46M	3
22	Kashmiri	6.55M	35	49	Kangri	1.12M	3
23	Bundeli	5.63M	35	50	Garo	1.13M	3
24	Mewari	4.21M	28	51	Kachchhi	1.03M	5
25	Awadhi	3.85M	36	52	Mahasu Pahari	1.00M	3
26	Wagdi	3.39M	35	53	Sanskrit	-	34
27	Lambadi	3.28M	28	54	Kodava	-	3
				55	Indian English	250M	43

Performance Tiers

Tier	BLEU Range	Count	Languages
Strong	35-43	26	Hindi, Bengali, Marathi, Telugu, Tamil, Gujarati, Urdu, Bhojpuri, Kannada, Malayalam, Odia, Punjabi, Rajasthani, Assamese, Maithili, Magahi, Marwari, Kashmiri, Bundeli, Awadhi, Wagdi, Nepali, Sindhi, Garhwali, Brajbhasha, Indian English
Good	32-34	7	Chhattisgarhi, Khortha, Bagheli, Kumauni, Ahirani, Banjari, Sanskrit
Adequate	20-28	9	Haryanvi, Mewari, Lambadi, Bhili, Harauti, Pahari, Sambalpuri, Nimadi, Surgujia
Partial	5-15	3	Konkani, Bagri, Kachchhi
Experimental	2-4	10	Dogri, Kurukh, Tulu, Manipuri, Santali, Kangri, Mahasu Pahari, Kodava, Bodo, Garo

Language Families

Indo-Aryan	Dravidian	Austroasiatic	Sino-Tibetan	Indo-European
43 languages	7 languages	1 language	3 languages	1 language

Model Architecture

Field	Value
Architecture	Encoder-Decoder (M2M-style)
Parameters	~1.3B (dense)
Encoder Layers	24
Decoder Layers	24
Model Dimension	1024
Attention Heads	16
FFN Dimension	8192
Activation	ReLU
Vocab Size	256,206
Tokenizer	SentencePiece BPE
Max Input Length	512 tokens
Max Positions	1024
Dropout	0.1
Languages	55 Indian
Translation Pairs	2,970
Scripts Supported	15+
CT2 Format	CTranslate2 (model.bin)

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

Use Cases

Government - Citizen services in every mother tongue | Sovereign deployment, data stays in India | Healthcare, education, judiciary outreach

Developers and Startups - Zero API cost for open-weight models | Build voice-first apps for any language | Fine-tune for domain-specific use cases | 2,970 translation pairs out of the box

Researchers and Academia - Full model weights for research | Benchmark against global state of art | Extend to more Indian languages | Advance Indian NLP and speech science

Limitations

• 10 languages are at experimental performance levels (BLEU 2-4): Dogri, Kurukh, Tulu, Manipuri, Santali, Kangri, Mahasu Pahari, Kodava, Bodo, Garo
• 3 languages have partial coverage (BLEU 5-15): Konkani, Bagri, Kachchhi
• Maximum input length is 512 tokens
• BLEU scores are tentative and based on human evaluation rather than standardized test sets

Citation

@misc{vaktranslate2026,
  title={Vak Translate: Open-Weight Translation for Every Indian Language},
  author={Sourav Bandyopadhyay, Ayush Kumar Bar, Aditya Singh Rathore, Grusha G S, Moksh K},
  year={2026},
  publisher={Shunya Labs},
  url={https://huggingface.co/shunyalabs/vak-translate-1.3b}
}

Contact

Shunya Labs

• Email: 0@shunyalabs.ai
• Web: shunyalabs.ai

---

Built in India. Open for everyone.