AI Today: Apertus 1.5 and NVIDIA in Industry

State of knowledge as of 2026-10-20. Source material gathered on 2026-10-20.

Today's overview covers key updates in the artificial intelligence ecosystem. We analyze the release of the Apertus 1.5 models, NVIDIA's strategic partnership with the German industrial sector, and the development of sovereign media data standards supported by the European Broadcasting Union. Additionally, we look at OpenAI's new initiatives regarding niche programming languages and the practical application of the Gemini model in professional outfit styling, pointing to the growing adoption of multimodal models across various economic sectors.

Key Takeaways (TL;DR)

• Apertus 1.5: Models 8B and 70B have been released with native image and audio support and a 262k token window.
• German Industrial AI Strategy with NVIDIA: A cooperation declaration was signed to advance AI in the industrial sector, including investments in German startups.
• NVIDIA and European Broadcasting Union Advance Media Data Sovereignty Standards: EBU and NVIDIA are implementing data sovereignty frameworks by integrating open AI models with European infrastructure.
• OpenAI Supports Niche Programming Languages: Evaluation sets have been published to improve ChatGPT's performance in SQLite environments and during COBOL code migration.
• Gemini Utilization in Professional Outfit Styling: The Gemini model supports outfit visualization and clothing care instruction generation in a professional stylist's workflow.

Apertus 1.5

According to an announcement published on October 15, the new version of Apertus 1.5 models has been released in 8B and 70B variants. These models continue the work on the first generation, offering native image and audio understanding, an optional thinking mode, a four times larger context window, and improved tool handling.

The update includes an expanded training dataset: 4 trillion tokens were added for the 8B model, and another 2 trillion for the 70B model, totaling 19T and 17T tokens respectively. An early multimodal data fusion approach was applied, where images (tokenized using Emu3.5) and audio (processed by WavTokenizer) enter the model as discrete tokens within the text sequence. The context window has been expanded to 262,144 tokens.

The training process included four preliminary stages and five fine-tuning stages (including SFT, RL, DPO, and alignment with the Apertus Charter). Model weights were released on Hugging Face under the Apache 2.0 license. The creators emphasize evaluation transparency by publishing reports in the Every Eval Ever (EEE) format, facilitating result verification in accordance with the European Union Artificial Intelligence Act (EU AI Act). In tests, the 8B model demonstrates high efficiency in multilingual tasks and translations, outperforming some models of similar scale.

For AI builders: Teams deploying multimodal models can test model performance in tasks based on documents, diagrams, and audio files. The information does not specify usage limits in production infrastructure or the computational resource cost required to handle the full 262k token context window. It is worth checking the real costs of handling multimodal tokens and the availability of technical documentation regarding specific commercial applications.

German Industrial AI Strategy with NVIDIA

German Chancellor Friedrich Merz and NVIDIA CEO Jensen Huang met on October 19 in Berlin to sign a cooperation declaration for the development of industrial artificial intelligence applications. The federal government aims to combine German engineering and manufacturing expertise with advanced AI technology to build a modern industrial ecosystem in Europe.

The main goal of the initiative is to deploy AI in the industrial sector, with a particular focus on small and medium-sized enterprises, known as the Mittelstand. Jensen Huang emphasized that the technology is meant to complement the skills of German engineers rather than replace them. The cooperation will be carried out within the "Industry 4.0" platform, connecting policymakers, academia, and industry representatives.

During the meeting, new NVIDIA investments in four German startups were announced: Agile Robots (industrial robotics), Avelios Medical (healthcare IT systems), Mirelo (audio and music generation), and Reflex Aerospace (satellite technologies). These investments are part of NVIDIA's nearly 10 billion dollar commitment supporting companies and research institutions in Europe.

Additionally, Siemens showcased AI solutions at its Erlangen factory that, according to company data, reduce engineering efforts in virtual production preparation by 85% and lower costs through agent-based planning by 20%. Simultaneously, a simulation of a 104-qubit quantum computer was conducted at the Jülich Supercomputing Centre using the JUPITER supercomputer.

For AI builders: Manufacturing companies can verify the integration capabilities of production management system data with AI models for "physical artificial intelligence". The announcement does not provide licensing terms for Siemens agentic solutions or implementation cost details for smaller Mittelstand entities – this needs to be checked before starting a pilot.

NVIDIA and European Broadcasting Union Advance Media Data Sovereignty Standards

During an industry event in Berlin on October 19, NVIDIA, Adobe, Bertelsmann, and UFA showcased progress in smart technologies for the media sector. The European Broadcasting Union (EBU) partnered with NVIDIA to expand data sovereignty frameworks for public broadcasters through cloud utilization, open AI models, and system integrator services.

This initiative defines requirements and reference architectures for data management, running computational workloads, and ensuring system security throughout their lifecycle. Cloud providers such as Microsoft and Nebius are involved in the project, offering solutions that allow sensitive editorial materials to be processed on European infrastructure. NVIDIA provides open Nemotron models that broadcasters can customize to their needs, languages, and editorial policies.

Parallelly, Adobe announced enhancements in the Frame.io tool, which uses NVIDIA GPU-accelerated computing to speed up video asset searches. Thanks to the use of the CUDA platform, TensorRT, and NVDEC, the embedding generation process is up to 5x faster, throughput has increased up to 4x, and search latency has decreased by 5x. German producer UFA is among those using these solutions, allowing creators to perform semantic searches for specific shots in libraries dating back 15 years. A new proxy format 80% smaller than the 360p version was also introduced, helping reduce required storage space.

For AI builders: A film or television producer can check whether their current video archives comply with the technical requirements for the semantic search model in Frame.io. Before deploying a sovereign infrastructure based on EBU and NVIDIA, it is worth verifying the maintenance costs of custom Nemotron models on European infrastructure and checking whether the chosen cloud provider ensures full compliance with local regulatory frameworks regarding editorial data. The source does not specify subscription costs or hardware requirements for local servers.

OpenAI Supports Niche Programming Languages

According to an announcement on October 19, OpenAI released new evaluation sets for software engineering research. These efforts focus on less popular programming languages, SQLite databases, and challenges related to COBOL code migration. The goal of the work is to improve ChatGPT's performance in specific programming environments.

For AI builders: A legacy systems architect can verify model behavior in the context of automating COBOL script migrations. The announcement does not specify whether support covers all language dialects, what the level of precision is in generating queries for SQLite, or whether these tools are available within the standard API – this needs to be checked before deployment.

Gemini Utilization in Professional Outfit Styling

According to information from October 19, 2026, stylist Sydney Bordonaro uses the Gemini model to create and visualize outfits for her clients, including professional athletes. The technology allows for building complete clothing sets remotely, eliminating the need for the stylist's physical presence during fittings.

The workflow with Gemini covers four main areas:

• Outfit Visualization: The stylist uploads clothing photos and client photos, and the model combines them into a single image. This allows checking how specific wardrobe elements look on a given person before making a purchase.
• Real-Time Modifications: Instead of creating new prompts, the user can request to swap individual clothing items, such as changing footwear or jacket styles, while keeping the rest of the outfit components intact.
• Decision Support: Ready visualizations are used for consultations with clients, facilitating decision-making regarding clothing choices for specific occasions, such as interviews or travel.
• Stain Removal: The model acts as an advisor in emergency situations – after uploading a photo of stained clothing, Gemini generates cleaning instructions based on locally available products.

For AI builders: A specialist dealing with supporting shopping processes can check the effectiveness of combining product images with a client's appearance. The publication does not specify the accuracy of fabric texture reproduction in generated images, real-time photo processing limits, or the availability of the clothing cleaning feature for various fabric types and stains, which requires verification before deployment in a production environment.

Frequently Asked Questions (FAQ)

What new features were introduced in the Apertus 1.5 models?

Apertus 1.5 models offer native image and audio understanding, an optional thinking mode, and a four times larger context window. The training dataset was increased to 19T tokens for the 8B model and 17T for the 70B model, utilizing an early multimodal data fusion method with Emu3.5 and WavTokenizer tokenizers.

What is NVIDIA's partnership with the German Mittelstand about?

The partnership aims to deploy AI in small and medium-sized enterprises as a tool to support engineers' competencies. NVIDIA is investing 10 billion dollars in European innovations, and companies like Siemens are deploying agent-based solutions that reduce engineering efforts by 85 percent.

What Gemini features does stylist Sydney Bordonaro use in her work?

Gemini enables outfit visualization by combining clothing photos with a client's appearance, making real-time modifications, and supporting purchasing decisions. The model also acts as an advisor, generating clothing cleaning instructions based on stain photos sent by the user.

Sources

• Apertus 1.5 Release Notes — 2026-10-15
• Germany Wants to Lead the World in Industrial AI: How NVIDIA Will Help — 2026-10-19
• NVIDIA Advances Agentic Security for Media, From European Public Service Broadcasting to Global Content Production — 2026-10-19
• Advancing coding in low‑resource programming languages — 2026-10-19
• 4 tips for styling outfits with Gemini — 2026-10-19