ashtok897/oroZ-1.7B-Indic-v1

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 6, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

oroZ-1.7B-Indic-v1 by ashtok897 is a 1.7 billion parameter base language model, pretrained from scratch on the Qwen3 architecture with a context length of 2,048 tokens. It is specifically designed for 13 Indic languages and English, utilizing a retrained tokenizer to significantly improve tokenization efficiency for Indic scripts. This model's primary differentiator is its Indic language focus and optimized tokenization, making it suitable for research and development in multilingual NLP for the Indian subcontinent.

Loading preview...

Overview

oroZ-1.7B-Indic-v1 is a 1.7 billion parameter base language model developed by ashtok897, built upon the Qwen3 architecture. Unlike many models, it was pretrained from scratch (not fine-tuned from existing weights) using a specialized dataset, ashtok897/indic-hplt-v2, which comprises 13 Indic languages and English. A key innovation is its retrained tokenizer, which drastically reduces the number of tokens per word for Indic languages (e.g., Hindi: 4.58 to 2.25, Telugu: 15.25 to 5.25 compared to the base Qwen3 tokenizer), leading to more efficient processing.

Key Capabilities

  • Multilingual Support: Designed for 13 Indic languages (Hindi, Bengali, Telugu, Marathi, Tamil, Urdu, Kannada, Malayalam, Gujarati, Nepali, Punjabi, Odia, Assamese) and English.
  • Efficient Indic Tokenization: Features a custom Byte-level BPE tokenizer (152,012 tokens) that significantly improves tokenization efficiency for Indic scripts, reducing token count by up to 55% compared to the base Qwen3 tokenizer for the training split.
  • Base Model: Provides foundational language understanding and generation capabilities, suitable for further fine-tuning.

Good For

  • Research in Indic NLP: Ideal for researchers and developers focusing on natural language processing for the Indian subcontinent.
  • Custom Fine-tuning: Serves as a strong base model for instruction-tuning or safety alignment for specific applications in Indic languages.
  • Exploring Efficient Multilingual Tokenization: Useful for understanding the impact of specialized tokenizers on language model performance in low-resource or complex script environments.

Finetunes
1 models