Neckarsulm/Rostock, 24 June 2026 – Planet AI, the Bechtle Group’s AI research and development company, has won the Competition on Multimodal Reasoning over Documents 2026 (DocVQA 2026). Held as part of the International Conference on Document Analysis and Recognition (ICDAR), the competition is widely regarded as the world’s most demanding benchmark for document understanding AI systems. The Rostock-based company secured first place both overall and in seven of the eight competition categories, outperforming the standard mixture-of-experts baseline by around 20 percentage points and the second-placed team by just under 14 percentage points.
Jesper Kleinjohann, CEO of Planet AI: “Winning DocVQA 2026 was not an end in itself, but it demonstrates that European AI research can compete internationally while meeting essential standards, particularly in the public sector—data sovereignty, GDPR compliance, and genuine freedom to choose your infrastructure. As the Bechtle Group’s AI lab, we combine this with the implementation expertise companies and public authorities need for real-world use.”
Hosted by the Computer Vision Center (CVC) at the Universitat Autònoma de Barcelona, this year’s competition required deep reasoning, planning, and multi-step inference across real-world documents in eight distinct categories—business reports, scientific papers, slides, posters, maps, comics, infographics, engineering drawings. In the open 35-billion-parameter category, teams were free to use any architecture—from individual frontier models and mixture-of-experts setups to complex agentic flows, all built on the most advanced foundation models from Google, Anthropic, OpenAI, and Alibaba.
Multi-layer architecture behind the success
Planet AI’s Distributed Cognitive Architecture (DCA) coordinates multiple foundation models as a collaborative system, adding capabilities that individual models do not provide on their own. The architecture operates in three layers, with IDA—Planet AI’s award-winning Intelligent Document Analysis—providing deterministically structured text as the foundation. Four vision-language models—Gemini 3.1 Pro, Gemini 2.5 Pro, Claude Sonnet 4, and Qwen3.5—then process the document independently. Claude Opus 4.6 serves as the orchestrating reasoning agent, comparing outputs across models, identifying inconsistencies, weighting sources by document type, and querying individual reader agents before reaching a decision. The system is fully model-agnostic and benefits automatically from advances in foundation models, while supporting deployment in both on-premises and cloud environments.
Large foundation models have impressive knowledge of the world and strong language capabilities, but reasoning and intelligence require more—memory, executive control, and the ability to converge on reliable outcomes. DCA provides exactly these capabilities, extending foundation models into effective knowledge workers and opening up the potential for a distinct new product category,” says Welf Wustlich, CTO and founder of Planet AI.
World-class European AI expertise
Founded in Rostock in 2015, Planet AI has grown over more than a decade through collaborations with European universities and participation in research programmes such as FP7 and Horizon 2020. Its success at DocVQA 2026 also reflects the SPOC-AI (Single Point of Contact for AI) research project, which Planet AI has been running since August 2024 together with technology partner Altow Digital Innovation and the University of Rostock, with funding from the state of Mecklenburg-Western Pomerania through the Technology-Beratungs-Institut (TBI). This first top-tier result has been achieved more than a year ahead of the project’s scheduled completion in July 2027, demonstrating that publicly funded AI research in Germany is gaining international visibility.
The official award ceremony will take place at ICDAR 2026 in Vienna between 30 August and 4 September 2026. The scientific foundations are documented in the DocVQA 2026 technical report, available on Zenodo (Cern): https://zenodo.org/records/20819848