---
title: "AI for Nepal — Tarka"
description: "HimalayaAI's Nepali language, OCR and speech models on one OpenAI-compatible API, and hosting on Tarka for other models built for Nepal."
url: "https://tarkahq.com/ai-for-nepal/"
---

AI for Nepal

# Models built for Nepal, served by Tarka.

Call HimalayaAI’s Nepali language, OCR and speech models through one OpenAI-compatible API. If you’re building a model for Nepal, Tarka can host and serve it alongside them.

[Get an API key](https://tarkahq.com/console) [Host your model](https://tarkahq.com/ai-for-nepal/#host-your-model)

Nepali models Live

- `himalaya-q8`
- `himalaya-gemma-4-q8`
- `glm-ocr-nepali`
- `nepali-parler-tts`
- `whisper-nepali-medium`

Call them at `https://tarka.rest/v1` with a Tarka API key.

## Start with HimalayaAI’s models.

HimalayaAI is a research lab that trains open-source Nepali AI models and publishes Nepali datasets and benchmarks. Four of its models run on Tarka, next to a community model for Nepali speech recognition.

| Published by | Model | Type | Model IDs | Source |
| --- | --- | --- | --- | --- |
| HimalayaAI | [HimalayaGPT 0.5B](https://tarkahq.com/inference/himalayagpt-0-5b/)<br>A compact, Nepali-first instruction model. | Language model | `himalaya-q4` `himalaya-q8` `himalaya-bf16` | [Hugging Face](https://huggingface.co/himalaya-ai/himalayagpt-0.5b-it-gguf) |
| HimalayaAI | [Himalaya Gemma 4 E2B](https://tarkahq.com/inference/himalaya-gemma-4/)<br>A larger Nepali–English instruction model based on Gemma 4 E2B. | Language model | `himalaya-gemma-4-q4` `himalaya-gemma-4-q8` `himalaya-gemma-4-bf16` | [Hugging Face](https://huggingface.co/himalaya-ai/himalaya-gemma-4-e2b-it-gguf) |
| HimalayaAI | [GLM OCR Devanagari](https://tarkahq.com/utilities/glm-ocr/)<br>GLM-OCR fine-tuned to read text from Devanagari documents. | Document OCR | `glm-ocr-nepali` | [Hugging Face](https://huggingface.co/himalaya-ai/glm-ocr-devanagari-finetuned) |
| HimalayaAI | [Nepali Parler TTS](https://tarkahq.com/utilities/nepali-parler-tts/)<br>Indic Parler-TTS fine-tuned for Nepali. Describe the voice you want in words. | Text to speech | `nepali-parler-tts` | [Hugging Face](https://huggingface.co/himalaya-ai/indic-parler-tts-nepali-finetuned-dgx-v2.0-cosine) |
| lilgoose777 | [Whisper Medium Nepali](https://tarkahq.com/utilities/whisper-nepali-medium/)<br>Whisper Medium for Nepali transcription, as a CTranslate2 build. | Speech to text | `whisper-nepali-medium` | [Hugging Face](https://huggingface.co/lilgoose777/whisper-medium-ne-ct2) |

Tarka keeps the inputs and outputs of HimalayaGPT and Himalaya Gemma requests for evaluation and debugging, not for training. Their model IDs ending in `-private` skip this; each model’s page lists them.

[Visit HimalayaAI](https://www.himalayaai.org/) [HimalayaAI on Hugging Face](https://huggingface.co/himalaya-ai)

## Make your first request.

Each example is the OpenAI Python client pointed at Tarka. Only the method and the model ID change.

1. ### Create an account

   Sign up in the console and request HimalayaAI models and Utility models.
2. ### Create an API key

   Give it the same model categories. It is shown once, so store it somewhere safe.
3. ### Call a model

   Use a model ID from this page. `GET /v1/models` lists the ones your key can call.

```python
from openai import OpenAI
 
client = OpenAI(
    base_url="https://tarka.rest/v1",
    api_key="tk_live_xxx",
)
prompt = "नेपालबारे कविता लेख्नुहोस्।"
chat = client.chat.completions.create(
    model="himalaya-q8",
    messages=[{"role": "user",
               "content": prompt}],
)
print(chat.choices[0].message.content)
```

```python
from openai import OpenAI
 
client = OpenAI(
    base_url="https://tarka.rest/v1",
    api_key="tk_live_xxx",
)
 
with open("voice-note.wav", "rb") as audio:
    result = client.audio.transcriptions.create(
        model="whisper-nepali-medium",
        file=audio, language="ne",
    )
print(result.text)
```

```python
from openai import OpenAI
 
client = OpenAI(
    base_url="https://tarka.rest/v1",
    api_key="tk_live_xxx",
)
 
speech = client.audio.speech.create(
    model="nepali-parler-tts",
    voice="A clear, calm Nepali speaker",
    input="नमस्ते नेपाल", response_format="wav",
)
speech.write_to_file("namaste.wav")
```

[Open the console](https://tarkahq.com/console) [Read the API docs](https://tarkahq.com/docs/inference/api-access/)

## Host your Nepali model on Tarka.

If you’ve trained or fine-tuned a Nepali language, speech or document model and its license allows hosting, Tarka can serve it through the same API as the models above.

1. **Send us the model** A Hugging Face repository or the weights, the license they’re released under, and a few example inputs with the outputs you expect.
2. **Tarka sets it up and tests it** Tarka’s engineers deploy the weights and check the output against your examples before it goes live.
3. **Developers call it** It gets a model ID on the Tarka API and a row in the table above, credited to you. Any account with access to that type of model can call it.

## Tell us about your model.

Send a link to the model and a line on what it’s for.

[Email hello@tarkahq.com](mailto:hello@tarkahq.com?subject=Hosting%20a%20Nepali%20model) [Create an account](https://tarkahq.com/console?auth=signup)