NVIDIA 擴大面向歐洲主權 AI 的開放模型組合

NVIDIA on October 21, 2026, announced an expansion of its family of open models spanning language, robotics, biology and quantum computing, a portfolio the company is positioning around Europe’s sovereign intelligence. According to NVIDIA, the open models give organizations a foundation they can customize with local data and deploy on infrastructure they control, allowing developers, researchers and enterprises to build applications tailored to local languages and industries.

The portfolio includes the new NVIDIA Nemotron 3.5 Super VL and Nemotron 3.5 Bolt models; new NVIDIA Nemotron Retriever models for embedding, reranking and document parsing; the new NVIDIA Kumo family of open-weight foundation models and libraries for structured business data; NVIDIA Cosmos 3.5 world models; NVIDIA Alpamayo 2 Edge for autonomous vehicles; NVIDIA BioNeMo for healthcare and scientific discovery; and NVIDIA Ising for quantum computing.

Nemotron Ecosystem and Enterprise Adoption

The Nemotron line supplies open language models together with datasets and training recipes for developers building agents. According to NVIDIA, the models handle both text and images and are intended to help agents coordinate workflows, work through complex problems and carry out specialized tasks at a lower cost. The models draw on data, research and expertise contributed through the NVIDIA Nemotron Coalition, including European partners Black Forest Labs, H-Company and Mistral.

Dassault Systèmes, Domyn, Dream, customers of Palantir, Salesforce, SAP, Shopify, Synopsys and WPP, as well as professional services providers Accenture and EY, are working with Nemotron models and datasets to create custom models and build domain-specific agents, according to NVIDIA. The company also named a post-training partner ecosystem customizing Nemotron for domain-specific workloads: AfterQuery, AgileRL, AI21, Alice, Applied Compute, Baseten, CoreWeave, Distil Labs, EON, Fastino, Fireworks, Google Cloud, Mercor, Multiverse Computing, Nebius, Handshake AI, Plurai, Prime Intellect, River.AI, Scale AI, Trajectory and TNG.

The portfolio expansion follows NVIDIA’s October 15, 2026, release of Nemotron 3.5 Super VL and the Nemotron 3.5 Bolt edge models, which added a 120-billion-parameter vision-language model and two locally deployable Bolt variants to the family.

Kumo Models for Structured Enterprise Data

The Kumo line consists of open foundation models, libraries and tools aimed at structured business data, where they can be used to predict outcomes, flag fraud risk and suggest next actions for agents and applications. As part of NVIDIA Agent Knowledge Foundation, Kumo is intended to shorten time to prediction and remove the need to build a separate machine learning pipeline for each task, reducing cost and complexity, NVIDIA says.

The Kumo portfolio includes NVIDIA Kumo Structured, with models for relational and tabular prediction, forecasting and anomaly detection, and NVIDIA Kumo Engine, an open library stack for developing and deploying structured-data foundation models. Developers can start with pretrained Kumo models and then fine-tune them on proprietary data when higher performance is critical.

Cognite, Coinbase, Databricks, DoorDash, Revolut and Tubi are among the first to work with Kumo models to accelerate development, NVIDIA said, and Snowflake customers are already using the models in industries including advertising, energy, financial services and retail.

Cosmos Tools and Alpamayo 2 Edge

Cosmos, NVIDIA’s family of open world foundation models, targets developers building robots, autonomous vehicles and vision agents that operate in the physical world. The new Cosmos post-training and optimization tools are meant to improve world models with less compute, and new simulation checkpoints let developers evaluate behavior in interactive driving and robotic manipulation scenarios before real-world testing.

NVIDIA also introduced Cosmos Data Factory, a set of blueprints and core agent skills for curating, labeling, generating and evaluating training data covering driving scenarios, product defects, image augmentation and event generation. Developers can run the full workflows or pick individual capabilities, with agents coordinating the process. Porsche with Voxel51, ANYbotics with Encord, Zalando, CEVA Logistics with Sereact, and Siemens are using Cosmos Data Factory to create training datasets and improve model precision, according to NVIDIA.

For autonomous driving, NVIDIA Alpamayo 2 Edge packages real-time scene understanding and motion planning in a compact reasoning model that runs directly in the vehicle. The model builds on the Nemotron-based foundation of Cosmos 3 Edge and is optimized for the NVIDIA DRIVE AGX Thor platform; NVIDIA says it lets developers deliver advanced driving features with lower in-vehicle computing requirements. Alpamayo 2 Edge will be available for download on GitHub and Hugging Face later in October 2026.

Protein Design and Quantum Models

NVIDIA BioNeMo RFD4-Proteina is an open protein design model developed by NVIDIA and the Baker Lab at the University of Washington’s Institute for Protein Design. It generates protein sequences and three-dimensional structures, supporting research into proteins that bind difficult targets, assemble into complex structures or attach to small molecules. NVIDIA reports that RFD4-Proteina delivers leading performance across more than 10 protein design tasks and generates designs about 13 times faster than RFdiffusion3 in a benchmark of 200-amino-acid proteins. Through agents, researchers can generate and evaluate candidate designs against design goals and molecular constraints before selecting which ones to test in the lab.

NVIDIA Ising is a family of open models intended as a common, extensible foundation for quantum computing work, spanning real-time hardware calibration through error correction across different qubit modalities. IQM Quantum Computers is using Ising to calibrate its quantum processors, work that NVIDIA says shows specialized open calibration models outperforming large general-purpose models. SEEQC is pairing the NVIDIA NVQLink platform and Ising models with its Digital Link to satisfy the low-latency demands of real-time quantum error correction. AWS is integrating the Ising family into AWS Parallel Computing Service so researchers can build on the models ahead of deployment, NVIDIA said, noting that decoding is typically performed on classical hardware colocated with the quantum processor because of latency requirements.

Availability

Nemotron 3.5 Super VL is available on Hugging Face, ModelScope, OpenRouter and build.nvidia.com as an NVIDIA NIM microservice, as well as through NVIDIA cloud, post-training and inference partners. The Nemotron 3.5 Bolt E4B 4GB model is expected to be available on October 23, 2026.

Nemotron Retriever open models are available on Hugging Face, various partner platforms and as an NVIDIA NIM microservice on build.nvidia.com. Kumo open models are available on Hugging Face and packaged as an NVIDIA NIM microservice for production inference on build.nvidia.com. Cosmos is openly available on Hugging Face and GitHub and through select inference partners.

The BioNeMo RFD4-Proteina open model is available on build.nvidia.com and as an NVIDIA NIM microservice, as well as on GitHub, Hugging Face and through select inference partners. Ising is available on GitHub, Hugging Face and build.nvidia.com.