NVIDIA Releases Nemotron and Cosmos Open Models at Scale

NVIDIA opened 2026 by releasing a coordinated set of open models spanning five domains: language, physical AI, autonomous vehicles, robotics, and biomedical research. The headline numbers are substantial: 10 trillion language training tokens, 500,000 robotics trajectories, 455,000 protein structures, and 100 terabytes of vehicle sensor data, all released alongside the model weights themselves.

The release clusters around two main platforms. The Nemotron family covers agentic language tasks through three specialized variants: speech, multimodal retrieval-augmented generation (RAG), and safety filtering. Cosmos handles physical AI, with the Alpamayo family targeting autonomous vehicle development, Isaac GR00T addressing robotics, and Clara serving biomedical applications.

For practitioners, the combination of weights and training data is what distinguishes this from a typical model drop. Open weights have become routine in 2025 and 2026. Open weights plus curated domain data at this scale is rarer, and it directly lowers the cost of fine-tuning for specialized tasks.

The industry adoption

The adopter list spans automotive, cybersecurity, enterprise software, and autonomous systems. Bosch is deploying Nemotron Speech for in-vehicle voice interaction. ServiceNow is training its own Apriel model family on open Nemotron datasets, citing cost efficiency for multimodal workloads. Cadence and IBM are piloting Nemotron RAG to improve search and reasoning over dense technical documents.

On the security side, CrowdStrike, Cohesity, and Fortinet are integrating Nemotron Safety models to layer trustworthiness into their AI pipelines. Palantir is embedding Nemotron into its Ontology framework to build an integrated stack for specialized artificial intelligence agents. CodeRabbit, an AI code review platform, is using Nemotron to scale review throughput. Uber and Hitachi are also listed as adopters, though NVIDIA has not published specifics on those deployments.

Broadening beyond hardware, this is deliberate strategy. By releasing open models that major enterprise software companies build on top of, NVIDIA creates a dependency chain that supports its broader ecosystem without requiring direct licensing revenue from the models themselves.

Gaps in the technical case

NVIDIA's announcement is notably light on benchmark comparisons. The speech, RAG, and safety variants each target narrow tasks where general benchmarks like GPQA or MMMU are largely irrelevant. The metrics that matter for multimodal RAG, such as recall at scale, latency under load, and hallucination rates on domain-specific documents, are not provided.

Sites like LLM Stats track releases and performance across the frontier, but specialized domain models often fall outside standard leaderboard coverage. Independent evaluation is necessary before committing any of these models to production use.

Physical AI claims are harder still to verify externally. The 500,000 robotics trajectories cited for Cosmos training is a large number, but trajectory count alone says nothing about diversity or edge-case coverage. Whether Isaac GR00T generalizes outside NVIDIA's reference hardware setups remains an open question practitioners will need to answer for themselves.

Broader context

This release continues a pattern that has been building since late 2024: hardware companies moving up the stack into model development and open data curation. NVIDIA's advantage over competitors making similar moves is the scale of its developer ecosystem and the ability to validate model performance on its own infrastructure before shipping.

For the artificial intelligence research community, the training data contribution may matter more than the models over the long run. Assembling 455,000 protein structures or 100 terabytes of vehicle sensor data requires institutional resources most labs cannot fund. Making that data openly available changes the economics of domain-specific research in ways that are harder to measure but potentially more durable.

Trackers like AI Release Tracker and Price Per Token will watch whether Nemotron variants appear on hosted inference APIs at competitive pricing, which would signal confidence in commercial viability beyond current enterprise partnerships.

If NVIDIA sustains this release cadence through 2026, the open-weight landscape will look substantially different by year-end. The real test is whether enterprise adoption translates into performance gains that justify the dependency on NVIDIA's model stack.

---

FAQ

Q: What is NVIDIA Nemotron?
A: Nemotron is NVIDIA's open model family for agentic AI, with variants covering speech processing, multimodal retrieval-augmented generation, and safety filtering for enterprise deployments.

Q: What is NVIDIA Cosmos?
A: Cosmos is NVIDIA's open platform for physical AI, encompassing the Alpamayo models for autonomous vehicles, Isaac GR00T for robotics, and Clara for biomedical research applications.

Q: How much open training data did NVIDIA release alongside these models?
A: The release includes 10 trillion language training tokens, 500,000 robotics trajectories, 455,000 protein structures, and 100 terabytes of vehicle sensor data.

Q: Which companies are currently deploying NVIDIA's open models?
A: Confirmed adopters include Bosch, ServiceNow, Cadence, IBM, CrowdStrike, Cohesity, Fortinet, Palantir, CodeRabbit, Uber, Hitachi, Salesforce, and Franka Robotics.