datalab-to / image-text-to-text updated 2 months ago

surya-ocr-2

Name: surya-ocr-2
Author: datalab-to

- Accuracy - scores 83.3% on olmOCR-bench (top under 3B params) - Speed - throughput of 5 pages/s on an RTX 5090 - Multilingual - scores 87.2% on an internal benchmark set of 91 languages (more here) - Layout analysis (table, image, header, etc.) with reading order - Table recognition (rows + columns)

Params

690 M

Context

—

Downloads 30d

1.2 M

Likes

Commercial use: conditional · openrail Not gated SAFETENSORS View on Hugging Face ↗

Download history

daily snapshots · 10 days

1.2 M873 K

Jul 28Jul 31Aug 3Aug 6

Can you run it?

Estimated VRAM at 8K context unless noted. Pick your hardware to see the verdict per quantization.

File	Quant	Size	Est. VRAM	Verdict on RTX 4090 · 24 GB
model.safetensors	bf16	1.4 GB	2.1 GB	✅ Runs comfortably

Estimate: file size × 1.1 + KV cache at 8K + 0.5 GB overhead. Not a benchmark — how we calculate this.

Specifications

Architecture: Qwen3_5ForConditionalGeneration
Parameters: 690 M
Tensor type: BF16
Licence: openrail
First seen on the Hub: 2026-05-14
Training datasets: undisclosed
Added to our catalog: 2026-07-28

Compare with

gemma-4-31B-it 11.9 M gemma-4-26B-A4B-it 11.7 M Qwen2.5-VL-7B-Instruct 9.3 M Qwen3.6-35B-A3B-FP8 8.9 M Qwen3.6-27B-FP8 7.8 M Qwen3.6-27B 7.0 M

Appears in

Best open coding models Commercial-use text models Models that fit in 12 GB