ProductsResearchSolutionsAboutCareersContact
Start a Project
← Benchmarks/Vision · Document AI

TinyDoc-VLM-256M vs. Donut: CPU Document AI Benchmark

Receipt OCR, Invoice Field Extraction, and Table Layout Analysis without GPUs. Evaluation of document understanding throughput, memory overhead, and OCR accuracy between TinyDoc-VLM-256M and Donut across CPU enterprise servers.

EXECUTIVE SUMMARY & EMPIRICAL VERDICT

TinyDoc-VLM-256M processes document pages 3.6× faster on CPU than Donut while delivering higher extraction fidelity on complex tabular and invoice data.

EULOGIK EDGE MODELFeatured

TinyDoc-VLM

256M (SigLIP + SmolLM2) · 256 Million
Target HardwareCommodity CPU / Laptop / Edge Node
Latency / Speed<180ms per document page
RAM / Memory512 MB RAM
Fertility / SizeONNX CPU optimized
  • ✓ Modern SigLIP vision backbone handles dense typography
  • ✓ SmolLM2 text decoder parses tabular structures directly
  • ✓ Runs directly on CPU with ONNX Runtime
  • ✓ Zero GPU requirement for high-volume invoice processing
Download Open Weights on HuggingFace ↗
INDUSTRY BASELINEOpen Weights

Donut VLM

base (Swin Transformer) · 140 Million
Target HardwareGPU recommended; slow on CPU
Latency / Speed>650ms per document page (CPU)
RAM / Memory1.1 GB RAM
Fertility / SizePyTorch native
  • • Pioneered OCR-free document parsing
  • • Prone to hallucination on small font contracts
  • • Substantially slower inference on standard CPUs
  • • Requires PyTorch runtime environment
METRIC SPECIFICATION MATRIX
METRICTINYDOC-VLMDONUT VLM
CPU Latency per Page<180 ms ★>650 ms
Invoice Key Extraction F191.8% ★84.2%
Dense Table Layout ParsingSupported ★Degrades on complex grid
Memory Overhead512 MB ★1.1 GB
ONNX Runtime CPU BinaryPre-compiled ★Complex export required
REPRODUCIBLE CODE SNIPPET

Execute this benchmark locally

from PIL import Image
from transformers import AutoProcessor, AutoModelForVision2Seq

processor = AutoProcessor.from_pretrained("eulogik/TinyDoc-VLM-256M")
model = AutoModelForVision2Seq.from_pretrained("eulogik/TinyDoc-VLM-256M")

image = Image.open("invoice.png")
inputs = processor(images=image, text="Extract total amount and tax", return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=128)
print(processor.decode(output[0], skip_special_tokens=True))

Interested in deploying TinyDoc-VLM?

We assist enterprises with custom fine-tuning, hardware sizing, and on-premise air-gapped installation.