NVIDIA Goes All-In on Open Source with Cosmos 3 and Nemotron Ultra
Published on 02.06.2026
NVIDIA Goes All-In on Open Source with Cosmos 3 and Nemotron Ultra
TLDR: NVIDIA dropped Cosmos 3, a fully open omnimodal world model family targeting physical AI and robotics, plus Nemotron Ultra at 550B parameters, and announced the RTX Spark personal AI superchip. The open-source angle is not marketing: the weights, code, datasets, and fine-tuning recipes all shipped.
Summary:
NVIDIA has been a hardware company that tolerated open source. This week feels different. Cosmos 3 ships with weights, code, datasets, and fine-tuning recipes, which is a full-stack open release in a way that CUDA ecosystem projects rarely have been. That matters, and I want to be specific about why.
Cosmos 3 uses a Mixture-of-Transformers architecture that pairs an autoregressive reasoner with a diffusion generator. The Nano configuration is 16 billion parameters (8B reasoner plus 8B generator), and the Super configuration is 64 billion parameters split evenly. Text-to-image and image-to-video fine-tunes from this family are now the top open-weight models in their categories. NVIDIA also launched the Cosmos Coalition with Runway as a partner, which is ecosystem building done seriously.
The physical AI framing is where it gets interesting. NVIDIA is positioning Cosmos 3 as infrastructure for robotics and physical automation, not just another multimodal chat model. Unifying language, image, video, audio, and action in a single architecture is aimed at the robot-that-needs-to-understand-its-environment problem. If that framing holds up, this is genuinely different from what the language model labs are doing.
Nemotron Ultra is 550 billion total parameters with 55 billion active (a mixture-of-experts model), announced by Jensen Huang at Computex in Taiwan. Multiple people in the AI space called it the strongest US open-weight model available. The combination of scale and efficiency is worth noting. A 550B MoE that is practically fast is a different kind of artifact than a dense 70B.
The RTX Spark personal superchip is previewed but not shipping yet. One petaflop on a personal computer is a statement about where NVIDIA thinks local inference is going. The partnership with Microsoft and the Hermes Agent launch partner suggest this is being positioned for developers running agentic workloads locally, not just gamers.
Key takeaways:
- Cosmos 3 is a genuinely open release (weights, code, data, recipes) using a Mixture-of-Transformers architecture targeting physical AI and robotics, not just text or code generation
- Nemotron Ultra at 550B parameters (55B active via MoE) is being called the strongest open-weight US model, combining scale with practical inference speed
- RTX Spark signals NVIDIA's intent to make one petaflop of personal AI compute a consumer reality, with agentic developer workloads as a clear target use case
Why do I care: The open-weight competition is accelerating faster than most product roadmaps can track. Two things stand out to me. First, Cosmos 3 being fully open including training data and recipes means the research community can actually build on it, not just run inference. That is a multiplier effect that proprietary releases do not get. Second, the physical AI framing is a reminder that the next wave of AI applications is not more chatbots, it is machines that need to perceive and act in the physical world. If NVIDIA has a head start on open infrastructure for that, the implications for who builds what over the next few years are significant. Meanwhile, the "40% of PRs include AI-generated code" stat floating around this week is a separate signal worth watching: the developer workflow is already changing faster than most teams are acknowledging.
Open Weights, Physical World
Understand why NVIDIA's fully open Cosmos 3 release is a bigger deal for robotics and physical AI than anything the language model labs shipped this week.