[Models](/models)/[Document Analysis](/models?category=document)/ibm-granite/granite-docling-258M

HFDocument StructureApache-2.0

# granite-docling-258M

by ibm-granite

Ultra-compact 258M document converter: layout, tables, formulas, and code in a single model

150Kdl/month

258Mparams

[HuggingFace](https://huggingface.co/ibm-granite/granite-docling-258M)[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)

Identifiers

Model ID

`ibm-granite/granite-docling-258M`

## Deploy granite-docling-258M

Single-tenant

Mixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.

[Read the customer-hosted guide](https://mixpeek.com/docs/customer-hosted/overview) [Talk to us about single-tenant](/demo)

## Overview

Granite-Docling-258M is IBM's ultra-compact vision-language model for end-to-end document conversion to machine-readable formats. Built on the Idefics3 architecture with a SigLIP2-base-patch16-512 vision encoder and a Granite 165M language model, it converts document pages into DocTags: IBM's universal markup format that captures all page elements including charts, tables, forms, code, equations, footnotes, and their spatial relationships.

At just 258M parameters, Granite-Docling rivals systems several times its size on layout detection (mAP 0.27), full-page OCR (F1 0.84), table recognition (TEDS 0.96), and equation recognition (F1 0.968). On Mixpeek, it provides the most cost-effective document structure extraction, converting scanned PDFs, contracts, and technical documents into structured, searchable content with full layout preservation.

## Architecture

Idefics3 architecture with SigLIP2-base-patch16-512 vision encoder and Granite 165M LLM. Outputs DocTags markup format describing all page elements and their spatial relationships. Experimental support for Japanese, Arabic, and Chinese. English is the primary target language.

## Mixpeek SDK Integration

IngestRetrieve

// No extractor parameter takes a Hugging Face model id (checked against
// GET /v1/discovery/extractors, which returns 13), so granite-docling-258M runs
// on your side and the output is upserted through POST
// /v1/namespaces/{namespace\_id}/documents/upsert. On Enterprise the other
// path is to upload the weights instead: POST /v1/namespaces/{id}/models
// accepts the huggingface format and a custom plugin loads them.
const res = await fetch(
  "https://api.mixpeek.com/v1/namespaces/ns\_your\_namespace/documents/upsert",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer API\_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      collection\_id: "col\_your\_collection",
      documents: \[
        {
          document\_id: "asset-00412",
          // The model produces text, so it lands in payload. Give the
          // collection a text vector index and embed that text to make it
          // searchable rather than only filterable.
          payload: { extracted\_text: modelOutput, source\_key: "archive/2026/asset-00412" },
          vectors: { "text-embedding": embeddingOfModelOutput },
        },
      \],
    }),
  },
);

// Managed alternative, if this exact model is not the requirement:
// document\_graph\_extractor@v1 runs intfloat/multilingual-e5-large-instruct
// (1024-d) over a bucket, with no inference of your own.

## Capabilities

*   Layout-preserving document-to-markup conversion
*   Table recognition (TEDS 0.96 on FinTabNet)
*   Equation recognition (F1 0.968) and code recognition (F1 0.988)
*   DocTags universal format for structured output
*   Ultra-compact 258M parameters: 3x smaller than SmolDocling

## Use Cases on Mixpeek

Contract processing: extract clauses, tables, and structured data while preserving layout

Technical documentation: convert manuals with code blocks, equations, and diagrams

Financial document extraction: parse statements, reports, and filings into structured data

## Benchmarks

Dataset

Metric

Score

Source

FinTabNet (table recognition)

TEDS (structure + content)

0.96

IBM, 2025: Granite-Docling Announcement

Full-page OCR

F1

0.84

IBM, 2025: Granite-Docling Announcement

Equation recognition

F1

0.968

IBM, 2025: Granite-Docling Announcement

## Performance

Input SizeFull document page (512px patches)

GPU Latency~65ms / page (A100)

GPU Throughput~15 pages/sec (A100, batch 4)

GPU Memory~0.7 GB

## Common Pipeline Companions

[

BAAI/bge-m3

Embed extracted text sections for semantic search

](/model/BAAI/bge-m3)[

docling-project/SmolDocling-256M-preview

Compare document structure extraction approaches

](/model/docling-project/SmolDocling-256M-preview)

## Explore on Mixpeek

[

More Document Analysis models

Compare alternatives in this category

](/models?category=document)[

Best Document AI Platforms

Hand-picked tools & platforms compared

](/curated-lists/best-document-ai-platforms)[

OCR & Document AI Internals

Deep-dive technical guide

](/guides/ocr-document-ai-internals)[

Feature Extractors

See how Mixpeek runs models as extractors

](/extractors)[

Mixpeek Vector Store

Store & search embeddings at scale

](/mvs)[

Pricing

Usage-based pricing for pipelines

](/pricing)[

All Curated Lists

Compare models, APIs & infrastructure

](/curated-lists)

### Specification

FrameworkHF

Organizationibm-granite

FeatureDocument Structure

Outputstructure tokens

Modalitiesdocument

RetrieverSection Filter

Parameters258M

LicenseApache-2.0

Downloads/mo150K

[

### Research Paper

Granite-Docling: End-to-End Document Understanding

arxiv.org](https://www.ibm.com/new/announcements/granite-docling-end-to-end-document-conversion)

### Build a pipeline with granite-docling-258M

Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.

[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)

### Alternative Models

[

microsoft/layoutlmv3-base

Document Structure

](/model/microsoft/layoutlmv3-base)[

naver-clova-ix/donut-base

Document Structure

](/model/naver-clova-ix/donut-base)[

fastino/gliner2-base-v1

Document Structure

](/model/fastino/gliner2-base-v1)[

docling-project/SmolDocling-256M-preview

Document Structure

](/model/docling-project/SmolDocling-256M-preview)

### Related in Document Analysis

[

microsoft/table-transformer-detection

Table Extraction

](/model/microsoft/table-transformer-detection)[

microsoft/table-transformer-structure-recognition-v1.1-all

Table Extraction

](/model/microsoft/table-transformer-structure-recognition-v1.1-all)[

poloclub/UniTable

Table Extraction

](/model/poloclub/UniTable)[

microsoft/table-transformer-structure-recognition-v1.1-pub

Table Extraction

](/model/microsoft/table-transformer-structure-recognition-v1.1-pub)