hariharasudhan/en-indic-translate-26b

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 11, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

The hariharasudhan/en-indic-translate-26b model is a 26 billion parameter Mixture-of-Experts (MoE) model, fine-tuned from google/gemma-4-26B-A4B-it, specifically designed for English to 11 Indic language translation. It excels at translating complex scientific documents, preserving LaTeX formulas, code blocks, and document structure across languages like Hindi, Bengali, Tamil, and Telugu. This model is optimized for maintaining document integrity and coherence over long, structured inputs, significantly outperforming baselines in math preservation and structural integrity.

Loading preview...

Overview

The hariharasudhan/en-indic-translate-26b model is a specialized English-to-Indic language translation model, built upon the google/gemma-4-26B-A4B-it base model. This is a Mixture-of-Experts (MoE) architecture with 26 billion total parameters (4 billion active). Its core strength lies in translating complex, structured English documents into 11 Indic languages: Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, and Telugu.

Key Capabilities

  • Preserves Document Structure: Accurately translates prose while leaving LaTeX math, section headings, and fenced code blocks untouched.
  • Handles Long, Complex Inputs: Designed to maintain coherence and avoid repetition loops over thousands of tokens in densely structured documents.
  • High Fidelity Translation: Achieves superior performance in math preservation (0.980), structure integrity (0.991), and renderability (98.2%) compared to baselines like Sarvam-Translate.
  • Low Degeneracy Rate: Exhibits a significantly lower rate of repetition loops (1.8%) and shorter loop periods, ensuring more usable output.
  • Reference-Free Evaluation: Performance is measured against source documents for structural and content integrity, not human-translated gold standards.

Good For

  • Translating scientific papers, technical manuals, or academic texts from English to supported Indic languages.
  • Use cases requiring precise preservation of mathematical equations and code snippets during translation.
  • Applications where maintaining the original document's layout and structural elements (headings, code blocks) is critical.
  • Developers needing a robust solution for high-quality, structured English-to-Indic document translation.

Finetunes
1 models