amad-iq/amad-vlm6

VISIONPricing:Input $0.4 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:32kPublished:Sep 9, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

amad-vlm6 is a 7 billion parameter vision-language model developed by amad-iq, created by merging two Arabic OCR fine-tunes of Qwen2.5-VL-7B-Instruct using the TIES method. This model specializes in Arabic Optical Character Recognition (OCR), particularly for document transcription, and is optimized to reduce 'thinking' behavior and improve output discipline compared to its parent models. It offers improved character error rate (CER) and character F-score (CHrF) on the KITAB-Bench `ocr-eval` benchmark, making it suitable for Arabic document processing tasks.

Loading preview...

amad-vlm6 Overview

amad-vlm6 is a 7 billion parameter vision-language model developed by amad-iq, specifically designed for Arabic Optical Character Recognition (OCR). It is a TIES merge of two Arabic OCR fine-tunes of Qwen2.5-VL-7B-Instruct: amad-iq/amad-vlm5 and AhmedZaky1/DIMI-Arabic-OCR-V2. This model aims to improve transcription accuracy and output discipline for Arabic documents.

Key Capabilities

  • Enhanced Arabic OCR: Achieves a macro Character Error Rate (CER) of 0.235 and a macro Character F-score (CHrF) of 81.68 on the KITAB-Bench ocr-eval benchmark across 13 datasets.
  • Improved Output Discipline: The merge significantly reduces 'thinking' rates (from 45-98% to 5-18%) and eliminates truncated or empty outputs, leading to fewer insertions and deletions in transcriptions.
  • Behavioral Gains: While reading accuracy (substitution rate) may be worse on some individual datasets, the overall improvement comes from better output behavior rather than superior character perception.
  • Flexible Usage: Available in bf16 safetensors for Transformers and vLLM, and Q8_0 + mmproj GGUF format for llama.cpp and LM Studio.

Good For

  • Arabic Document Transcription: Ideal for tasks requiring accurate text extraction from Arabic images and documents.
  • Reducing Model Hallucinations: Specifically targets and mitigates certain types of over-generation and 'thinking' behaviors observed in its parent models.
  • Further Fine-tuning: Provides a strong base for specialized Arabic OCR applications.

Limitations

It's important to note that the model's reading accuracy (substitution rate) regresses on 10 of 13 datasets, with CER wins primarily from output discipline. There are also significant regressions on specific datasets like isippt and muharaf due to imported over-generation, making it less suitable for presentation-slide or muharaf-style handwritten inputs. The benchmark results are specific to transcription accuracy on 13 Arabic document styles and do not cover other aspects like mixed-language text, tables, or varying image quality.

Finetunes
3 models