ml-intern-explorers/gemma-4-e2b-arabic-english-vision

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:5.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Apr 28, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

ml-intern-explorers/gemma-4-e2b-arabic-english-vision is a 5.1 billion parameter vision-language model, pruned from google/gemma-4-e2b-it, specifically optimized for Arabic and English tasks. It retains full vision and text generation capabilities while reducing model size by 17% through the removal of unnecessary language tokens and the audio encoder. This model excels in bilingual Arabic/English vision-language tasks, document understanding, and serves as an efficient base for fine-tuning.

Loading preview...

Overview

This model, ml-intern-explorers/gemma-4-e2b-arabic-english-vision, is a specialized version of Google's gemma-4-e2b-it model. It has been strategically pruned to optimize performance and efficiency for Arabic and English vision-language tasks.

Key Optimizations

  • Vocabulary Reduction: The original 262,144-token vocabulary was reduced to 209,836 tokens, focusing on Arabic, English, and special tokens. This involved removing CJK, Cyrillic, Devanagari, and Thai language support.
  • Audio Encoder Removal: The 12-layer audio encoder present in the original model was removed.
  • Size Reduction: These pruning steps resulted in a 17% reduction in total model size, from 10.2 GB to 8.5 GB (bf16), primarily by shrinking the Per-Layer Embedding (PLE) tables.
  • Vision Capabilities Retained: The 16-layer vision encoder and the 35 decoder layers remain fully intact, ensuring no loss in image understanding or text generation capabilities for the targeted languages.

Performance

Benchmarking on the loay/ar_stage1_probe dataset for Arabic text recognition showed identical performance to the original google/gemma-4-e2b-it model, with a CER of 0.1168 and 23/100 exact matches. This demonstrates that the pruning is lossless for Arabic and English tasks.

Intended Use Cases

  • Bilingual Arabic + English vision-language applications.
  • Document understanding tasks involving both languages.
  • As a more efficient base model for fine-tuning on Arabic/English specific datasets.
  • Any scenario where support for CJK, Cyrillic, or Indic languages is not required.

Finetunes
1 models