OCR models / Alibaba (Qwen)

Qwen3-VL 4B

Name: Qwen3-VL 4B
Author: Alibaba (Qwen)

Alibaba (Qwen) · released 2025-10-11 · data checked 2026-11-06

Qwen3-VL 4B is a 4.4B-parameter general vision-language model from Alibaba (Qwen) (CN), released 2025-10-11 under the Apache-2.0 license. It scores 79.2 on olmOCR-Bench (#14 of 24). It needs about 9.9 GB of memory at 16-bit.

A general vision-language model, not an OCR specialist; listed for comparison.

OmniDocBench v1.6

—

Not reported

olmOCR-Bench

79.2

#14 of 24 models

Parameters

4.4B

General vision-language model

Memory to run

~9.9 GB

16-bit weights plus about 1 GB

Benchmarks

olmOCR-Bench Third party79.2

Measured by CodeSOTA. Best: OXR 1.0 87.8

The tick marks the best score on each benchmark. See how to read the ranking.

Specifications

Lab: Alibaba (Qwen) (CN)
Type: General vision-language model
Total parameters: 4.4B
License: Apache-2.0
Memory to run: ~9.9 GB at 16-bit
Released: 2025-10-11

Weights and usage on Hugging Face ↗

Closest alternatives

dots.ocr (3B, 79.1 on olmOCR-Bench, MIT)
Qianfan-OCR (4.7B, 79.8 on olmOCR-Bench, Apache-2.0)
DeepSeek-OCR 2 (3.4B, 76.3 on olmOCR-Bench, Apache-2.0)
Falcon OCR 1.5 (270M, 82.4 on olmOCR-Bench, Apache-2.0)

Sources

Something wrong here? Report it.