Rostock, June 22nd 2026 – PLANET AI GmbH, the AI research and development company of the Bechtle Group, has won the ICDAR Competition on Multimodal Reasoning over Documents 2026 (DocVQA 2026). TThe Rostock-based company finished first not just overall, but in seven of the eight competition categories — with a lead of 22.5 percentage points over the best frontier-model baseline (37.5%) and 13.75 percentage points over the second-placed team (46.25%).

DocVQA 2026 is widely regarded as the world’s most demanding AI benchmark for document understanding. Organized by the Computer Vision Center (CVC) at the Universitat Autònoma de Barcelona, it requires deep reasoning, planning, and multi-step inference across real documents in eight entirely different categories: business reports, scientific papers, presentation slides, scientific posters, infographics, maps, comics, and engineering drawings. A system that excels in just one category fails in the others.

Architecture beats model size

In the open category (over 35 billion parameters), every team could compete with any architecture — from single frontier models to Mixture-of-Experts configurations to complex agentic flows, each built on the strongest foundation models from Google, Anthropic, OpenAI, and Alibaba. For PLANET AI, the decisive factor was not model size but architecture.

PLANET AI’s approach, the Distributed Cognitive Architecture (DCA), coordinates multiple foundation models as a cooperative team. It adds capabilities that individual models structurally lack: memory across long documents, an executive layer that detects contradictions, and a convergence dynamic that builds answers step by step rather than in a single pass.

The architecture operates in three layers. IDA, PLANET AI’s Intelligent Document Analysis engine, provides deterministically structured text as the foundation. The underlying technology has already won seven international ICDAR/ICFHR competitions between 2014 and 2019. DocVQA 2026 is the eighth. Four vision-language models (Gemini 3.1 Pro, Gemini 2.5 Pro, Claude Sonnet 4, Qwen3.5) read the document independently, each optimized for its own strengths. Claude Opus 4.6 acts as the orchestrating reasoning agent — comparing outputs, detecting conflicts, weighting sources by document type, and issuing targeted follow-up queries to individual reader agents in roughly 10 percent of cases before committing to a final answer. The system is fully model-agnostic and benefits automatically from advances in foundation models.

Welf Wustlich, CTO und Gründer von PLANET AI: „Large AI models — foundation models — are world models with impressive knowledge and language capabilities. But thinking and intelligence require more: memory, executive control, and a convergence dynamic. That is exactly what DCA adds, turning foundation models into intelligent knowledge workers. I believe this is the foundation for a paradigm shift. DocVQA 2026 is the first public proof that this path holds.“

A new product category: digital knowledge workers

The result has a clear implication for any organization running document-intensive processes — in finance, insurance, public administration, manufacturing, and beyond. A system that handles the full range from business reports and scientific posters to infographics, maps, comics, and engineering drawings handles invoices, contracts, and applications just as well. The challenges are the same: mixed layouts, text and visual content combined, information spread across multiple pages, tables with critical figures.

Classical document processing treats documents as data sources: fields are read, booked into systems, checked off. What PLANET AI delivers with DCA is something different — the ability to treat documents as knowledge objects: extract, relate, detect contradictions, reason across multiple steps, and answer follow-up questions. That is a product category of its own.

Jesper Kleinjohann, CEO von PLANET AI: „Winning DocVQA 2026 is not an end in itself. It demonstrates that European AI research is internationally competitive while meeting requirements that are non-negotiable — especially in the public sector: data sovereignty, GDPR compliance, and independence from US infrastructure. As the AI lab of the Bechtle Group, we can combine that with the implementation strength that enterprises and public authorities need for real-world deployment.“

Sovereign AI from Europe

PLANET AI was founded in Rostock in 2015 and has grown over more than a decade through collaborations with European universities and within European research programs including FP7 and Horizon 2020. The system runs entirely within the customer’s own infrastructure on request — GDPR-compliant, with no dependency on US cloud services. Sovereignty is part of the DNA, not a compliance layer added as an afterthought.

The victory is also an early result of the SPOC-AI research project, which PLANET AI has been running since 2024 in collaboration with ALTOW and the University of Rostock, funded by the State of Mecklenburg-Vorpommern through the Ministry of Economic Affairs. The Distributed Cognitive Architecture behind PLANET AI’s DocVQA 2026 win is a direct outcome of this work, more than a year ahead of the project’s official end date in July 2027.

The official award ceremony will take place at ICDAR 2026 in Vienna (August 30 – September 4, 2026). The scientific foundations are documented in the competition’s technical report: https://doi.org/10.5281/zenodo.20707289

About PLANET AI

PLANET AI pioneers automation for document-driven business processes using German AI technology. Its core solutions include IDA for high-precision document capture and JAIDE as an AI assistant. Together, they ensure end-to-end data quality from initial capture through decision-making. Since 2023, PLANET AI has been part of the Bechtle Group, serving scanning service providers, software manufacturers, and document-intensive industries ranging from insurance and banking to public administration. The solutions combine modular architecture and flexible deployment models with strict adherence to European data protection standards. Founded in Rostock in 2015, PLANET AI employs around 50 professionals.

Further links:

• ICDAR 2026: https://icdar2026.org/
• DocVQA 2026 Competition: https://rrc.cvc.uab.es/?ch=34&com=introduction
• Official winner announcement: https://rrc.cvc.uab.es/?ch=34&com=news&view=data&id=83
• Leaderboard: https://rrc.cvc.uab.es/?ch=34&com=evaluation&task=1
• Technical report (Zenodo): https://doi.org/10.5281/zenodo.20707289