Language model
HimalayaGPT 0.5B
Sovereign Nepali LLM by Himalaya AI Research Lab, hosted in Kathmandu.
- Status
- Available
- Parameters
- 0.5B
- Context
- 2K
- Best for
- Nepali-language inference — civic services, Nepali content, multilingual agents
Model IDs
| Model ID | Precision | Use |
|---|---|---|
himalaya-q4 | Q4 K M | Smallest footprint |
himalaya-q8 | Q8 | Balanced default |
himalaya-bf16 | BF16 | Highest-precision variant |
himalaya-q4-private | Q4 K M | Private · no internal retention |
himalaya-q8-private | Q8 | Private · no internal retention |
himalaya-bf16-private | BF16 | Private · no internal retention |
Use one of the model IDs above with Chat Completions or Responses at https://tarka.rest/v1.
Pricing
himalaya-q4: Free
himalaya-q8: Free
himalaya-bf16: Free
himalaya-q4-private: $0.03 input · $0.01 cached input · $0.06 output per million tokens
himalaya-q8-private: $0.04 input · $0.01 cached input · $0.08 output per million tokens
himalaya-bf16-private: $0.05 input · $0.01 cached input · $0.10 output per million tokens
Overview
HimalayaGPT 0.5B is a 500-million-parameter instruction-tuned model built by Himalaya AI Research Lab and trained on a Nepali-language corpus with a custom Nepali BPE tokenizer. The project’s stated mission is digital sovereignty for AI-native governance in Nepal — deployed in production behind civic services like the Nagarik App and Hello Sarkar.
Tarka hosts HimalayaGPT 0.5B on our Kathmandu infrastructure. If you’re building something Nepali — civic tech, a Nepali-language chatbot, a translation pipeline, or an offline-replacement experiment — point your OpenAI SDK at our endpoint.
Why it is in the catalog
We’re an AI and agent infrastructure company based in Kathmandu. Making the country’s sovereign LLM discoverable is an easy decision. The work that went into HimalayaGPT belongs to the community that built it.
Fair-use limits still apply (we will rate-limit obvious abuse — spam, resale, runaway loops) but real Nepali-language workloads will not hit them.
When to pick it
- Anything Nepali-language: civic tech, government workflows, content for Nepali audiences, Nepali chat agents
- Bilingual routing where the Nepali path needs a compact dedicated model
- Edge / offline replacement experiments where a 0.5B model on local hardware is the right tool
- Educational and research use — first-class welcome here
When to look elsewhere
- Larger Nepali–English workloads → use Himalaya Gemma 4 E2B
- Long-context reasoning beyond the deployed 2K route
- General coding agents that need a larger model and broader training mix
Credit
The model and weights are by the Himalaya AI Research Lab team. We are the host, not the author. The Hugging Face page and the lab’s GitHub have the model cards, tokenizer, and training pipeline.
Data retention
The original free aliases retain inputs and outputs in Tarka’s private internal archive for evaluation and debugging until Tarka administrators delete them. They are not used for model training.
Choose an alias ending in -private to avoid internal retention. Private aliases use the inference API-key scope and are currently priced at zero. Organization capture, when enabled, writes only to your organization’s own bucket.