GTC Berlin: NVIDIA expands its open model family

Updated: October 21, 2026, 13:45 (Budapest)

What changed

On October 21, the day of the GTC Berlin keynote, NVIDIA expanded its open model portfolio (NVIDIA blog). The package includes Nemotron 3.5 Super VL, released last week, so we skip it here. The new pieces:

• NVIDIA Kumo: open foundation models and libraries for structured business data. Forecasting, fraud risk scoring, recommendations and anomaly detection from tables, without a separate ML pipeline per task. It consists of the Kumo Structured models and the open Kumo Engine library.
• Nemotron Retriever: five new open models, also licensed for commercial use (Hugging Face). Nemotron 3.5 Embed 8B and 1B search across text and document images (tables, charts, infographics); the 1B is also available in an NVFP4-quantized version. Nemotron 3.5 Rerank 8B refines the ranking of results, while the 631-million-parameter Nemotron 3.5 Parse turns document images into structured text. According to NVIDIA's measurements, Embed 8B ranks first among dense retrieval models on the ViDoRe v3.1 leaderboard with an nDCG@10 score of 61.17, Rerank 8B ranks first on the reranking leaderboard with 60.49, and Parse ranks first among sub-1-billion-parameter models on ParseBench with 75.20. In an agentic test, the reranker cut per-query token cost by about 13-14%.
• Cosmos: new post-training and optimization tools for improving world models with less compute, simulation checkpoints for driving and robotic manipulation tests, plus Cosmos Data Factory blueprints for curating, labeling, generating and evaluating training data.
• Alpamayo 2 Edge: a compact reasoning model for real-time scene understanding and motion planning in autonomous vehicles, optimized for DRIVE AGX Thor.
• BioNeMo RFD4-Proteina: an open protein design model built with the Baker Lab at the University of Washington. NVIDIA says it leads on more than 10 design tasks and generates about 13x faster than RFdiffusion3 on 200 amino acid proteins.
• RFD4-Proteina is also available through partners including Anthropic's Claude Science, OpenAI, Google Cloud and Amazon Bio Discovery. NVIDIA says it beat evaluated baselines on design success rate across four benchmark suites (vendor data). Separately, Siemens Healthineers is exploring Nemotron models for agentic radiotherapy co-planning.
• Ising: an open model family for quantum processor calibration and error correction (NVIDIA Developer). AWS is integrating it into AWS Parallel Computing Service.

What ships now

Kumo, Nemotron Retriever, Cosmos, RFD4-Proteina and Ising are available on Hugging Face or GitHub, several also as NIM microservices on build.nvidia.com. Alpamayo 2 Edge follows later this month, and the Nemotron 3.5 Bolt E4B 4GB model is expected on October 23. Nemotron Retriever models also run as NIM microservices and on Baseten, DeepInfra and FriendliAI clouds.

Why it matters

Kumo is the most interesting new piece. Most company data lives in tables, where LLMs are weak. If a pretrained model plus fine-tuning can replace a hand-built ML pipeline per task, agents can ask for predictions straight from company data. The announcement gives no performance numbers, so test it on your own data before replacing anything.