DEV.co
Open-Source LLM · ibm-granite

granite-docling-258M

Granite Docling 258M is a lightweight multimodal model (258M parameters) designed to convert document images into structured text and markdown. Built by IBM Research, it combines a vision encoder (SigLIP2) with a small language model (Granite 165M) and integrates directly into the Docling document processing pipeline. It handles PDFs, images, and URLs, extracting text, tables, formulas, and layout information with support for English and experimental support for Japanese, Arabic, and Chinese.

Source: HuggingFace — huggingface.co/ibm-granite/granite-docling-258M
258M
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
70k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developeribm-granite
Parameters258M
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / taskimage-text-to-text
Gated on HuggingFaceNo
Downloads70k
Likes1.2k
Last updated2025-09-23
Sourceibm-granite/granite-docling-258M

What granite-docling-258M is

A Vision-Language Model (VLM) based on Idefics3 architecture with custom modifications: vision encoder replaced with siglip2-base-patch16-512 and language backbone with Granite 165M. Operates as an image-text-to-text model outputting DocTags (structured markup) compatible with docling-core for conversion to markdown, HTML, and other formats. Supports inference via transformers, vLLM, ONNX, and MLX-VLM frameworks. Trained on SynthCodeNet, SynthFormulaNet, SynthChartNet, and DoclingMatrix datasets. Released September 17, 2025.

Quickstart

Run granite-docling-258M locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="ibm-granite/granite-docling-258M")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

PDF to Markdown/HTML Conversion at Scale

Batch process research papers, technical documents, or enterprise PDFs into clean markdown or HTML using vLLM. Model size and integration with Docling pipeline enable cost-effective server deployment.

Document Parsing and Extraction Pipelines

Extract structured data (tables, formulas, text blocks) from scanned or digital documents for downstream processing, compliance workflows, or archival systems. DocTags output provides rich semantic information.

Research and Academic Document Processing

Parse academic papers (arxiv, journals) to extract equations, figures, tables, and text in a queryable format. Experimental multilingual support suits international research repositories.

Running & fine-tuning it

ESTIMATE: ~1–2 GB VRAM for inference in bfloat16 (typical GPU deployment). CPU inference feasible with MLX (Apple Silicon) or quantized ONNX. Batch inference via vLLM benefits from 8–24 GB VRAM for throughput. No quantization details in card; assume full precision baseline; ONNX/MLX variants may reduce footprint. Verify actual memory on your hardware before production.

Not clearly stated in card. Model is presented as a production inference tool integrated into Docling. No mention of LoRA, QLoRA, or domain-specific fine-tuning. Possible but requires custom training setup and reintegration with Docling pipeline. Not recommended unless addressing significant domain shift; prefer prompt engineering or RAG-based augmentation first.

When to avoid it — and what to weigh

  • Real-Time Ultra-Low-Latency Requirements — Even at 258M parameters, inference latency for full documents is non-trivial. Not suitable for sub-100ms response SLAs. Batch processing is the intended use pattern.
  • Non-Document Image Understanding — Model is specialized for document layout and structure. Performance on general scene understanding, face recognition, or non-textual imagery is not documented and likely poor.
  • Proprietary/Sensitive Data without Infrastructure Control — If you require air-gapped or on-premises inference with strict data residency, model must be self-hosted. HuggingFace or cloud API endpoints may violate compliance.
  • Multilingual Documents at Production Scale — Japanese, Arabic, and Chinese support are marked experimental. Expect degradation on non-English documents; test thoroughly before production deployment.

License & commercial use

Apache License 2.0 (permissive OSI-approved license). Allows commercial use, modification, and distribution provided the license and copyright notice are retained.

Apache 2.0 is a permissive open-source license that explicitly permits commercial use without royalty or special permission. You may use Granite Docling 258M in proprietary products, SaaS platforms, and internal enterprise systems. Ensure you: (1) include a copy of the Apache 2.0 license, (2) document material changes if modified, and (3) retain copyright and attribution notices. No liability assumed by IBM Research under the license. Verify compliance with your legal team if integrating into regulated systems (healthcare, finance).

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityLow
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Model processes document images and text; consider data sensitivity. Card does not describe input validation, adversarial robustness, or attack surface. Deployment over untrusted networks should use HTTPS/TLS. Self-hosting is recommended for regulated data (PHI, PII). Model outputs are deterministic (greedy decoding in examples) but could be exploited if user-controlled prompts are allowed; apply input sanitization. No security audit or disclosure policy mentioned; no known CVE data in card.

Alternatives to consider

Docling SmolDocling-256M (ds4sd/SmolDocling-256M-preview)

Slightly smaller predecessor with similar capability; may have different training data or stability profile. Check if available and compare inference cost/quality.

Llama 3.2 Vision (Meta)

Larger (11B) general-purpose VLM; better at open-ended vision tasks but overkill and slower for document-specific extraction. Higher cost, more flexibility.

Qwen VL-Chat or Claude 3.5 Vision (API)

Proprietary closed-source alternatives offering higher accuracy and broader understanding. Higher cost per request but no infrastructure burden; suit low-volume or specialized queries.

Software development agency

Ship granite-docling-258M with senior software developers

Start with the quick-start CLI or Python SDK. For production batch processing, set up vLLM and monitor inference latency on your document corpus. Ensure your infrastructure meets the 1–2 GB baseline VRAM and choose self-hosting for regulated data. Review the Docling project documentation and test on your document types before scaling.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

granite-docling-258M FAQ

Can I use Granite Docling 258M in a commercial product or SaaS service?
Yes. The Apache 2.0 license permits commercial use without restriction or royalty. Ensure you include the Apache 2.0 license text in your product's legal notices and retain copyright attribution. Verify with your legal team if you are subject to export controls or regulated industries (healthcare, defense).
What is the minimum VRAM required to run inference?
Approximately 1–2 GB for single-image inference in bfloat16 on GPU. Batch inference with vLLM can use 8–24 GB depending on batch size and sequence length. CPU inference is possible with quantized ONNX or MLX, but slower. Test on your target hardware; actual memory depends on image resolution and max_new_tokens.
Does this model support languages other than English?
Primary support is English. Japanese, Arabic, and Chinese support are marked experimental in the card. Thoroughly test on your non-English document corpus before production. Expect degradation compared to English performance.
Can I fine-tune this model on my own documents?
Not clearly documented in the card. Model is designed as a pre-trained inference tool. Fine-tuning is theoretically possible but requires custom training infrastructure and reintegration with Docling. Not recommended; try prompt engineering, RAG, or in-context examples first. Contact IBM Research if interested in a training variant.

Work with a software development agency

DEV.co helps companies turn open-source tools like granite-docling-258M into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your open-source llms stack.

Ready to Deploy Granite Docling?

Start with the quick-start CLI or Python SDK. For production batch processing, set up vLLM and monitor inference latency on your document corpus. Ensure your infrastructure meets the 1–2 GB baseline VRAM and choose self-hosting for regulated data. Review the Docling project documentation and test on your document types before scaling.