AI Today: NVIDIA Agent Toolkit and IBM Granite

State of knowledge as of 2026-10-22. Source material gathered between 2026-10-21 and 2026-10-22. Today's overview focuses on broad updates to the NVIDIA ecosystem, covering AI agent creation tools, a new open model offering for European enterprises, and the cuDSS library for engineering simulations. Furthermore, we examine the update to IBM's Granite speech model series and the debut of the AnyLanguageModel 1.0 library, which simplifies local and hosted model management for developers building Apple platform applications.

TL;DR

• NVIDIA Agent Toolkit: A new toolkit enables enterprises to build and customize AI agents using CUDA-X libraries and Nemotron models.
• NVIDIA models: The company introduced open Nemotron, Kumo, Cosmos, and BioNeMo models, supporting European digital sovereignty in data processing.
• NVIDIA cuDSS: The library solves sparse linear systems on GPUs, enabling up to a 10x speedup in engineering simulations.
• IBM Granite Speech: The updated speech model series offers improved transcription performance and edge device optimization while retaining the Apache 2.0 license.
• AnyLanguageModel 1.0: A new Apple system library provides a unified API for managing local and hosted models and supports the LanguageModel protocol.

NVIDIA Agent Toolkit

On October 21, NVIDIA expanded the open Agent Toolkit, used by enterprises to create, customize, and control artificial intelligence agents leveraging corporate data. The expanded kit includes the Agent Knowledge Foundation, which utilizes CUDA-X libraries such as cuDF, cuGraph, cuML, cuOpt, and cuVS. The new Auto-Ontology library transforms raw business data into concepts like customers, products, or orders, mapping relationships between them. The package also features Nemotron Retriever models and the Kumo family (Tabular, Relational, Time Series, and Forecast versions). Agent safety is ensured by the OpenShell runtime environment, while NemoClaw design patterns are used to build specialized systems. Entities such as Palantir, Databricks, Snowflake, Amazon Web Services, SAP, ServiceNow, and Zoom already use the tools. ServiceNow reported nearly a 25% reduction in latency while improving ranking quality, whereas SAP accelerated its reranker model performance almost threefold. For AI builders: Engineering and data science teams can leverage CUDA-X libraries and Kumo models to deploy agents based on their own databases without moving files to external systems. Before deciding on deployment, one should check computing infrastructure maintenance costs, hardware requirements for CUDA-X libraries, and whether Kumo models support specific data formats used within the organization. The documentation also does not specify license requirements for individual components of the package or availability of support in Polish.

NVIDIA Expands Open Model Portfolio for European Enterprises

According to an announcement on October 21, NVIDIA launched a range of new open models covering natural language processing, robotics, biology, and quantum computing. The goal of the initiative is to support European digital sovereignty by providing tools that organizations can customize using their own data and deploy on infrastructure they independently control. The new offering includes Nemotron 3.5 Super VL and Nemotron 3.5 Bolt models, as well as the Nemotron Retriever family designed for embedding, reranking, and document parsing tasks. For business data, the NVIDIA Kumo model family was prepared to facilitate outcome forecasting and risk assessment, alongside the NVIDIA Kumo Engine library. In robotics and autonomous systems, NVIDIA Cosmos 3.5 models were released alongside the Cosmos Data Factory toolkit and the compact NVIDIA Alpamayo 2 Edge model optimized for the NVIDIA DRIVE AGX Thor platform. The portfolio is complemented by the NVIDIA BioNeMo RFD4-Protein model, developed in collaboration with the University of Washington, used for designing protein sequences and structures, which tests show generates designs about 13x faster than RFdiffusion3. Meanwhile, the NVIDIA Ising model family aims to support quantum supercomputer work, including hardware calibration and error correction. Most of the mentioned solutions are already available on the Hugging Face platform, GitHub, and as NVIDIA NIM microservices. For AI builders: Data engineers designing RAG systems and teams building autonomous agents can test Nemotron and Kumo family models on their own datasets. The announcement does not provide detailed licensing terms regarding the commercial use of individual models or a full list of limitations in processing languages other than English — this must be verified in technical documentation before launching a production deployment.

NVIDIA Releases cuDSS Library

On October 21, NVIDIA announced the general availability of the cuDSS library as part of the CUDA-X suite. This tool shifts the process of solving sparse linear systems to graphics processing units (GPUs), aiming to accelerate engineering and scientific simulations. The solution supports calculations for complex designs such as advanced integrated circuits, aircraft, or industrial systems. Using cuDSS allows for a reduction in time required to obtain results and an increase in simulation throughput, enabling the analysis of larger and more detailed models. The library is designed to scale from single workstations to large computing clusters. The tool can be accessed via interfaces in C, C++, Python, and Fortran. Industry leaders already use the technology by integrating it with their own software. For instance, Applied Materials reported up to 10x faster sparse system solving in its Ginestra platform, while ZEISS SMT combined with COMSOL Multiphysics achieved a 4x speedup in plasma simulations for EUV lithography. Other entities such as Dassault Systèmes, Cadence, Siemens, and Synopsys also report measurable performance gains in their computing tools. For AI builders: Engineering teams implementing numerical simulations can check cuDSS integration with their existing libraries in C, C++, Python, or Fortran. The announcement does not specify license terms or detailed pricing for using the library in commercial projects — these data must be confirmed directly with the supplier.

IBM Updates Granite Speech Model Collection

On October 21, IBM announced an update to its Granite Speech model series, introducing the new granite-speech-5.0-2b model and refreshed versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. The new granite-speech-5.0-2b model supports transcription and translation in English, French, German, Spanish, Portuguese, and Japanese, as well as translation from English to Italian and Mandarin. It was trained on 390,000 hours of public audio recordings and synthetic data. Model tests yielded an average Word Error Rate (WER) of 5.06% for the English language. The 470m-turboctc version was optimized for speed on edge devices. Thanks to Conformer block compilation optimization, the model achieves a performance of 19,870 RTFx, representing about a 1.6x increase compared to the launch version. Meanwhile, the granite-speech-4.1-2b-plus model was fine-tuned to handle long recordings such as meetings, interviews, or support calls while retaining the previous architecture and prompt format. All models were released under the Apache 2.0 license with native support in Transformers, vLLM, llama.cpp, and mlx-audio environments for Apple Silicon processors. The manufacturer recommends combining these solutions with the Granite Guardian model for risk detection. For AI builders: Teams operating multilingual support centers or automated meeting transcription systems can test these models in existing audio processing pipelines. The announcement does not state whether the models offer support for the Polish language or what actual infrastructure maintenance costs are at high request scale — these aspects should be verified prior to production deployment.

AnyLanguageModel 1.0

According to information from October 20, 2026, the AnyLanguageModel 1.0 library was released as a replacement for the Apple Foundation Models framework. This tool was designed to allow developers to more easily manage local and hosted models in applications on Apple platforms. Alongside the update, swift-huggingface 1.0 and swift-xet 1.0 packages were released, supporting model downloads from the Hugging Face platform. The main premise of AnyLanguageModel is to provide a unified API so that the same code session can handle system models, open-weight models run via llama.cpp, MLX, and Core ML, as well as selected third-party provider services. Version 1.0 introduces support for models compliant with the LanguageModel protocol, including Core AI models (on OS 27 systems). The solution also runs in Linux environments. New features include tool calling support, response streaming, and token usage reporting for hosted models. Developers can use API key passing mechanisms via closures, enabling secure retrieval of short-lived tokens from their own servers. When switching providers, reasoning blocks are automatically skipped if the new provider does not support their recreation. For AI builders: An iOS mobile app developer can check whether their current infrastructure supporting Foundation Models requires migration when switching to this replacement. It is worth verifying the library's license and whether the availability of external inference service providers covers regions compliant with the project's legal requirements, which the source does not specify.

Frequently Asked Questions (FAQ)

What tangible benefits does the deployment of NVIDIA cuDSS bring?

Deploying the cuDSS library allows for significant acceleration of numerical calculations. Applied Materials achieved up to 10x faster solving of sparse linear systems with it, and ZEISS SMT combined with COMSOL Multiphysics achieved a 4x speedup in advanced plasma simulations used in lithography.

What languages does the new IBM granite-speech-5.0-2b model support?

The granite-speech-5.0-2b model supports transcription and translation in English, French, German, Spanish, Portuguese, and Japanese. Furthermore, it offers translation functions from English to Italian and Mandarin. It was trained on 390,000 hours of public audio recordings and synthetic data.

Sources

• NVIDIA Agent Toolkit Expands, Grounding Enterprise Agents in Company Data — 2026-10-21
• NVIDIA Open Models Advance Europe’s Sovereign Intelligence Across Industries — 2026-10-21
• NVIDIA cuDSS Now Generally Available to Accelerate Engineering and Scientific Simulations — 2026-10-21
• Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster — 2026-10-21
• AnyLanguageModel 1.0: One Year of Local Models on Apple Platforms — 2026-10-20 --- Text prepared automatically by the aipoc-lab (AI) laboratory based on publicly available sources listed above. Own summaries, not copies. Pre-human-edited version.