OPEN LINK ↗
// 59d agoMODEL RELEASE
IBM drops Granite-4.0-3B-Vision for document extraction
IBM has released Granite-4.0-3B-Vision, a specialized vision-language model (VLM) optimized for enterprise-grade document data extraction. Built as a LoRA adapter on top of the Granite-4.0-Micro base, it excels at converting complex charts, tables, and semantic key-value pairs into structured formats like JSON, CSV, and HTML.
// ANALYSIS
IBM's "Western Qwen" moment delivers a surgical tool for the least sexy but most valuable enterprise task: turning messy PDFs into structured data.
- –Unique LoRA adapter design on a hybrid Mamba-2/Transformer base slashes RAM usage by 70%, allowing a single 3B deployment to pivot between multimodal and text-only tasks.
- –Specialized task tags for Chart2CSV and Table2JSON move beyond simple OCR, offering semantic extraction that integrates directly into IBM's Docling pipeline.
- –ISO 42001 certification and Apache 2.0 licensing provide the "clean" provenance required by regulated industries that often skip less-governed open-weights.
- –Optimized for 8GB VRAM consumer hardware while supporting massive context windows, it represents a direct challenge to Alibaba’s small-model dominance.
// TAGS
ibm-granitevlmmultimodalopen-weightsdoclingenterprise-aidocument-extractiongranite-4.0-3b-vision
DISCOVERED
59d ago
2026-03-28
PUBLISHED
59d ago
2026-03-28
RELEVANCE
9/ 10
AUTHOR
jacek2023
