Hugging Face

Shunya Labs publishes open weights for several models on Hugging Face. Pull them down for local inference, fine-tuning, or air-gapped deployment in environments where the cloud API isn't an option.

The Shunya Labs organization

All open models live under huggingface.co/shunyalabs.

Model	Task	Params	License	Notes
`pingala-v1-universal`	ASR	0.8B	Shunya Labs RAIL-M	3.10% composite WER on the OpenASR leaderboard. 204 languages. API name: `zero-universal`.
`pingala-v1-en-verbatim`	ASR (English verbatim)	-	-	English-focused variant on Hugging Face.
`zero-stt-hinglish`	ASR (code-switch)	0.8B	openrail	First Hinglish ASR that handles code-switched conversations natively. API name: `zero-codeswitch`.
`vak-translate-1.3b-ct2`	Translation	1.3B	CC-BY-SA-4.0	55 Indian languages, 2,970 any-to-any pairs. BLEU 38.5 weighted average.

Cloud API names → Hugging Face

Production APIs use short model IDs. Map them to open weights or demos on Hugging Face:

API model	Product	Hugging Face
`zero-universal`	Zero STT Universal	`pingala-v1-universal` (weights)
`zero-indic`	Zero STT Indic	Zero STT Space (demo)
`zero-med`	Zero STT Med	Zero STT Med Space (demo)
`zero-codeswitch`	Zero STT CodeSwitch	`zero-stt-hinglish` (weights)
`zero-indic`	Zero TTS	TTS-Indic Space (demo)
`vak-translate-1.3b-ct2`	Vāķ Translate	`vak-translate-1.3b-ct2` (weights)

Demos on Hugging Face Spaces

Space	Task	Notes
`TTS-Indic`	Text-to-speech	Vāķ Text to Speech, try Indic synthesis in the browser (Gradio). API: `zero-indic`.
`Zero-STT-Shunya-Labs`	Speech-to-text	Multilingual transcription with diarization (demo).
`Zero_STT_Med_Shunya_Labs`	Medical STT	Clinical / medical transcription (demo). API: `zero-med`.
`vak-speech-to-text`	Speech-to-text	Vāķ Speech to Text, 55 Indian languages (demo).

Authentication

Hugging Face access is gated by a personal access token. Pass it via the HUGGINGFACE_HUB_TOKEN environment variable, the CLI login command, or programmatic login().

shell

# Option 1, env var
export HUGGINGFACE_HUB_TOKEN=hf_...

# Option 2, CLI login (interactive)
huggingface-cli login

# Option 3, programmatic
from huggingface_hub import login
login(token="hf_...")

Download model weights

shell

huggingface-cli login
huggingface-cli download shunyalabs/pingala-v1-universal --local-dir ./models/pingala
huggingface-cli download shunyalabs/vak-translate-1.3b-ct2 --local-dir ./models/vak
huggingface-cli download shunyalabs/zero-stt-hinglish --local-dir ./models/hinglish

Use the Pingala ASR model directly

Shunya publishes a thin Python wrapper around the universal ASR model on PyPI as pingala-shunya.

shell

pip install pingala-shunya

python

from pingala_shunya import PingalaTranscriber

# Loads the model from a local path (handy for air-gapped runs)
tx = PingalaTranscriber(model_path="./models/pingala")
segments = tx.transcribe("meeting.wav")
for s in segments:
    print(f"[{s.start:.2f} → {s.end:.2f}] {s.text}")

Use the Vāķ translation model

The Vāķ model ships in CTranslate2 format (vak-translate-1.3b-ct2). Run it with the ctranslate2 + transformers + sentencepiece stack.

shell

pip install ctranslate2 transformers sentencepiece

python

import ctranslate2
from transformers import NllbTokenizer

tokenizer = NllbTokenizer.from_pretrained("./models/vak")
translator = ctranslate2.Translator("./models/vak", device="cuda")  # or "cpu"

src_lang = "eng_Latn"
tgt_lang = "hin_Deva"
text = "How are you today?"

tokens = tokenizer.convert_ids_to_tokens(tokenizer.encode(text))
results = translator.translate_batch([tokens], target_prefix=[[tgt_lang]])
print(tokenizer.decode(tokenizer.convert_tokens_to_ids(results[0].hypotheses[0][1:])))

Air-gapped deployment

For environments without internet connectivity (defence, offline kiosks, mobile edge), pull the weights once on an internet-connected machine and transfer them across:

Run the huggingface-cli download commands above on a connected workstation.
Package the ./models directory as a signed tarball.
Verify checksums, transfer via signed media to the secure zone.
Load the model from the local path, no further network calls needed.

Production serving

For multi-tenant, streaming, horizontally-scaled serving in an air-gapped zone, the on-prem Triton package is easier than rolling your own. See Deployment overview or contact Shunya Labs ↗.

Licence reminder

pingala-v1-universal, Shunya Labs RAIL-M: free up to 10,000 hours/month. No redistribution, no derivatives. Attribution required when outputs are public.
zero-stt-hinglish, openrail: permissive with responsible-use restrictions.
vak-translate-1.3b-ct2, CC-BY-SA-4.0: share-alike with attribution.

Always check the model card on Hugging Face before shipping, the licence file there is authoritative.

Source: Shunya Labs Hugging Face organization page (huggingface.co/shunyalabs) and the per-model cards linked above. Air-gapped deployment steps reproduced from the Shunya Labs Deployment guidance.