NVIDIA Drops Open-Weight Nemotron 3.5 and Six AI Families for Enterprise

At GTC Berlin, NVIDIA unveiled an open-weight portfolio for teams that want to fine-tune and run AI models in their own environments. The release covers language and vision, retrieval, structured business data, robotics, autonomous vehicles, protein design and quantum computing. NVIDIA is distributing weights through services such as Hugging Face while offering NIM microservices, packaged inference services optimized for NVIDIA hardware.

European AI procurement increasingly emphasizes data residency, regional infrastructure and control over deployment. Downloadable weights and self-hosted inference support those requirements more directly than the closed flagship models from OpenAI and Anthropic. Open-weight refers to access to model parameters; source code, training data and usage rights remain specific to each model and license.

Seven families, one portfolio

Nemotron widens the agent stack

The flagship release, Nemotron 3.5 Super VL, is a vision-language model that accepts text and images. The smaller Nemotron 3.5 Bolt targets lower-cost language workloads. Together with the Retriever family, the models cover reasoning, visual understanding, embeddings, reranking and document parsing for agentic applications.

NVIDIA developed the models with data, research and technical contributions from the Nemotron Coalition. Its European participants include Black Forest Labs, H Company and Mistral, aligning the portfolio with regional efforts to build AI systems that can run on locally controlled infrastructure.

NVIDIA named Dassault Systèmes, Domyn, Dream, Salesforce, SAP, Shopify, Synopsys and WPP among the organizations working with Nemotron models or datasets. The company also cited customers of Palantir and professional services firms Accenture and EY. These references cover a mix of evaluation, customization and domain-specific agent development rather than a single deployment stage.

Kumo brings tables into the model

The Kumo family addresses relational data such as transactions, user events, account histories and inventory snapshots. Its models are designed to forecast outcomes, identify anomalies, estimate fraud risk and recommend actions without requiring a separately engineered machine-learning pipeline for every prediction task.

NVIDIA identified Cognite, Coinbase, Databricks, DoorDash, Revolut and Tubi as early Kumo users. It also said Snowflake customers are applying the models in advertising, energy, financial services and retail.

Kumo targets work often handled by bespoke pipelines built with methods such as XGBoost. Production use still requires controls for data leakage, label quality, backtesting, drift and task-specific validation, especially when predictions affect financial or operational decisions.

Cosmos connects simulation to the edge

Cosmos 3.5 adds post-training tools and simulation checkpoints for evaluating robot and vehicle behavior before physical testing. World models learn representations of environments and possible actions, allowing developers to generate or simulate scenarios that would be expensive, rare or dangerous to reproduce with hardware.

The new Cosmos Data Factory automates parts of data curation, labeling and generation for scenarios such as unusual driving events and manufacturing defects. NVIDIA cited Porsche working with Voxel51, ANYbotics with Encord, Zalando and CEVA Logistics with Sereact, and Siemens as users creating training datasets through the system.

Alpamayo 2 Edge brings scene understanding and motion planning onto autonomous vehicles. The compact reasoning model uses the Nemotron-based Cosmos 3 Edge foundation and is optimized for NVIDIA’s DRIVE AGX Thor platform, connecting the company’s model software with its automotive compute hardware.

Protein design and quantum control

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. NVIDIA reports leading results across more than 10 protein-design tasks and says the model generated designs about 13 times faster than RFdiffusion3 in a benchmark involving 200-amino-acid proteins.

Higher generation throughput can expand the number of candidates considered in generate-and-screen workflows. The benchmark does not establish biological validity, so research teams still need computational filtering, laboratory testing and independent comparisons on their target protein classes.

The Ising models focus on calibration and error-correction workflows for quantum computers. IQM Quantum Computers is using Ising to calibrate quantum processors, while SEEQC is combining Ising and NVQLink with its Digital Link system for real-time error correction. AWS plans to integrate the family with AWS Parallel Computing Service.

Why NVIDIA is bundling the stack

1. Open-weight distribution extends NVIDIA’s platform strategy. Teams that fine-tune and serve these models are likely to rely on CUDA, NVIDIA GPUs and NIM, expanding demand for the company’s software and hardware.
2. European deployment requirements shape the packaging. Downloadable weights and regional or on-premises inference fit procurement rules centered on data residency, operational control and customization.
3. Specialized workloads broaden NVIDIA’s reach. Structured data, simulation, protein design and quantum calibration remain less crowded than general-purpose language modeling.

From model card to production

• Check licenses: Confirm rights for commercial use, fine-tuning, redistribution and generated outputs for each model.
• Measure hardware needs: Review parameter counts, supported precision, context length, quantization options, memory requirements and expected concurrency.
• Benchmark the deployment path: Compare direct weight hosting with NIM on latency, throughput, cost and operational complexity.
• Validate on local data: Reproduce vendor benchmarks using representative inputs, failure cases and domain-specific metrics.
• Review governance: Document training-data provenance where available, retention policies, telemetry, security updates and model versioning.
• Test portability: Measure performance outside CUDA and NIM if future hardware flexibility is a requirement.

Engineering teams can now assemble retrieval, reasoning, structured prediction and world-model components from one NVIDIA portfolio, then choose direct weight hosting or packaged NIM inference. That approach can simplify integration around CUDA while concentrating more of the AI stack with one supplier, a tradeoff teams should include in architecture and procurement decisions.