isaacmg/qwen3-vl-8b-hebrew-v20a-merged

VISIONPricing:Input $0.727 / Output $5.405Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The isaacmg/qwen3-vl-8b-hebrew-v20a-merged model is an 8 billion parameter vision-language model based on Qwen3-VL-8B-Instruct, developed by isaacmg. It specializes in diplomatic transcription of handwritten Hebrew-script manuscripts (Hebrew, Aramaic, Judaeo-Arabic) from the Cairo Genizah, offering advanced text grounding capabilities like phrase-to-box and box-to-text. This model is optimized for historical document analysis and transcription, achieving strong performance on Genizah religious and documentary benchmarks with a 32768 token context length.

Loading preview...

Qwen3-VL-8B Hebrew v2.0a — Merged Weights

This model, isaacmg/qwen3-vl-8b-hebrew-v20a-merged, is an 8 billion parameter vision-language model based on Qwen3-VL-8B-Instruct, specifically fine-tuned for the transcription and analysis of handwritten Hebrew-script manuscripts. It represents the current best in its series (as of September 2026) for transcription benchmarks and is the first checkpoint to include text grounding capabilities.

Key Capabilities

  • Diplomatic Transcription: Accurately transcribes handwritten Hebrew, Aramaic, and Judaeo-Arabic texts from Cairo Genizah fragments and religious/Talmudic manuscript pages.
  • Text Grounding: Features advanced grounding functionalities including:
    • locate: Phrase-to-bounding box mapping.
    • read_box: Extracting text from a specified bounding box.
    • Page-level line detection with associated bounding boxes.
  • High Performance: Achieves strong benchmark results with a median aligned F1 of 0.849 on Genizah religious-140 and 0.875 on Frozen PGP-131 documentary benchmarks. Grounding performance shows a median IoU of 0.47 for locate and a median CER of 0.13 for read_box.
  • Optimized Image Processing: Ships with a specific image processor contract (min_pixels=6,500,000, max_pixels=7,000,000) to handle large manuscript images efficiently.

Good For

  • Historical Document Analysis: Ideal for researchers and institutions working with ancient Hebrew-script manuscripts, particularly those from the Cairo Genizah.
  • Automated Transcription: Automating the transcription of complex handwritten texts, reducing manual effort.
  • Text Localization: Identifying and extracting specific phrases or lines within manuscript images.
  • Digital Humanities Projects: Supporting projects that require detailed analysis and digitization of historical texts.

Finetunes
1 models