NVIDIA Unveils New Open Models, Data and Tools to Advance AI Across Every Industry

TL;DR

NVIDIA’s CES 2026 omnibus open-release announcement, bundling new open models, datasets, and tooling across five domains: Nemotron (agentic LLMs / ASR / RAG / safety), Cosmos (physical-AI world foundation models), Alpamayo (autonomous-vehicle reasoning VLA), Isaac GR00T (humanoid robotics VLA), and Clara (biomedical / protein / RNA). The bundle ships 10T language tokens, 500K robotics trajectories, 455K synthetic protein structures, and 100TB of vehicle sensor data on Hugging Face and GitHub. From a research-team perspective the relevant items are the world-foundation-model line (Cosmos Reason 2, Cosmos Predict 2.5, Cosmos Transfer 2.5) and the open VLA pair (Isaac GR00T N1.6, Alpamayo 1) — the rest is enterprise plumbing.

Key claims

NVIDIA is releasing Cosmos Reason 2 (a reasoning VLM for physical-world understanding), Cosmos Predict 2.5 and Cosmos Transfer 2.5 (synthetic-video generators across diverse environments) as open world-foundation models [§Physical AI].
Isaac GR00T N1.6 is an open reasoning vision-language-action (VLA) model for humanoid robots, using Cosmos Reason as its reasoning backbone for full-body control [§Physical AI].
Alpamayo 1 is claimed as the first open, large-scale reasoning VLA model for autonomous vehicles, paired with AlpaSim, an open-source closed-loop simulation framework [§Alpamayo].
The accompanying Physical AI Open Datasets include 1,700+ hours of driving data spanning a wide range of geographies and edge cases [§Alpamayo].
Nemotron Speech ships a 0.6B streaming ASR model claimed at 10× faster than peers on Daily/Modal benchmarks while topping the Open ASR leaderboard [§Nemotron].
Nemotron RAG ships embed + rerank VLMs for multimodal/multilingual document retrieval, and Nemotron Safety adds Llama-3.1-Nemotron-Safety-Guard-8B-v3 and a Nemotron PII (GLiNER-based) detector [§Nemotron].
Clara contributes four model families: La-Proteina (atom-level protein design), ReaSyn v2 (synthesis-aware drug design), KERMT (computational safety / drug–body interaction), and RNAPro (3D RNA structure prediction) [§Clara]; the protein-structure dataset is released alongside ([arXiv:2512.01976]).
A 455,000-structure synthetic protein dataset is released to train downstream structure models [§Clara].

Method

This is a product announcement, not a research paper, so there is no unified method. The release pattern is: model weights + open dataset + reference workflow / blueprint, distributed via Hugging Face, GitHub, and build.nvidia.com (with NIM microservice packaging for deployment). The two architectural patterns worth flagging are (i) Cosmos Reason being reused as a perception/reasoning backbone inside Isaac GR00T N1.6 (i.e. world-model-as-VLA-prior) and (ii) the Alpamayo 1 / AlpaSim pairing, which positions a reasoning VLA model against a closed-loop simulator rather than only an offline driving benchmark. The blog itself contains no architectural diagrams; for any specific sub-release the model card on Hugging Face is the actual source.

Results

The post quotes two leaderboard claims with citations: Nemotron Speech ASR at the top of the Hugging Face Open ASR leaderboard with a “10× faster” claim against in-class peers, and Cosmos Reason 2 topping a “physical reasoning leaderboard” (HF space facebook/physical_reasoning_leaderboard). Cosmos Predict 2.5 / Transfer 2.5 are described as “leading” on a Physical AI Bench leaderboard (shi-labs/physical-ai-bench-leaderboard). No numbers are quoted for Alpamayo 1, Isaac GR00T N1.6, or any of the Clara models — those sit behind their respective model cards and a separate Alpamayo press release.

Why it’s interesting

The signal here is structural, not numerical. NVIDIA is consolidating its open-source surface around two world-foundation-model lines (Cosmos for general physical AI, Alpamayo specifically for AV) and using them as the reasoning prior for the embodied VLAs (Isaac GR00T N1.6, Alpamayo 1) — i.e. world model as backbone, VLA as task-head, rather than training each VLA from scratch. That architectural pattern is the same one Luma would care about for any video-as-reasoning-substrate work, and Cosmos Reason 2 is the most directly relevant open checkpoint to track against in-house video understanding. Worth pulling the individual model cards if any of these become baselines.