Nemotron 3 Ultra Lands as a 550-Billion-Parameter Open-Weight Model Topping US Open Rankings
If you have been waiting for an open-weight model you can actually self-host without settling for second-tier quality, this is the most interesting release of the day. Nemotron 3 Ultra ships with 550 billion parameters (55 billion of them active) and Nvidia claims the top spot on US open-weights rankings with an Intelligence Index of 48 and throughput above 300 tokens per second. The framing is squarely agentic: it is tuned for reasoning, planning and multi-step tool use rather than chat.
For teams that cannot send data to a closed API, an open model at this level changes what is realistic to build in-house, and the high token rate matters if you are chaining many calls in an agent loop. It surfaced first through Nvidia's own keynote and blog at COMPUTEX before the wider press wrote it up.
Cosmos 3 Arrives as the First Open Physical AI Omnimodel for Robotics and Simulation
This one matters less if you build chatbots and a great deal if you touch robotics, simulation or any system that has to reason about the real world. Cosmos 3 is pitched as the first open physical AI omnimodel and ranks first across more than seven robotics benchmarks. The point of Cosmos is understanding and simulating physical environments accurately enough to train and validate robots and autonomous systems in software before they touch hardware.
Nvidia released the models immediately in higher-accuracy and faster lightweight variants, with a local edge version promised soon. For anyone evaluating physical AI stacks, having an open model rather than a closed simulator lowers the barrier to experimenting significantly. It came via Nvidia's official COMPUTEX channels.
The Vera CPU and Rubin-Generation Silicon Complete Nvidia's COMPUTEX Infrastructure Picture
Nvidia used the keynote to launch six new pieces of silicon, led by the Vera CPU (a processor built specifically for agentic AI and reinforcement learning workloads, claimed to be twice as efficient and 50 per cent faster than traditional x86 server CPUs). The Vera Rubin NVL72 rack pairs 36 Vera CPUs with 72 Rubin GPUs over sixth-generation NVLink, alongside new ConnectX-9, BlueField-4 and Spectrum-6 networking parts. None of this is something you buy directly, but it sets the price and availability of the cloud capacity you rent, so cheaper or faster frontier inference over the next year traces back to days like this.
MiniMax M3 Launches as an Open-Weight Frontier Model with a 1 Million-Token Context Window
If your work leans on coding models or long-context tasks, M3 deserves a quick evaluation. MiniMax launched M3 today, positioning it as the first open-weight model to combine frontier-level coding, a one million token context window and native multimodal input in a single release. The architecture rests on a sparse attention mechanism that lets it hold that context without the usual cost blow-up, and early benchmark claims put it in the conversation with GPT-5.5 and Gemini 3.1 Pro.
Weights are expected on Hugging Face and GitHub within about ten days, which means self-hosting and fine-tuning are realistic shortly. For anyone weighing open alternatives to closed coding APIs, a credible open model with this much context is a meaningful new option rather than a curiosity.
Industry Themes
Two patterns stand out today. First, the open-weight tier is where the action is: both genuinely new models in the window (Nemotron 3 Ultra and MiniMax M3) are open rather than closed APIs, which continues a clear shift toward credible self-hostable options at or near frontier quality. For teams that have been treating open weights as a cost-cutting measure rather than a capability play, both of these warrant re-evaluation.
Second, the centre of gravity moved from the usual model labs to the layers above and below them, with Nvidia setting the hardware agenda at COMPUTEX while the established chat providers stayed quiet. The Nvidia news surfaced through the company's own keynote and blog before mainstream coverage caught up, so a feed that watched only general tech press would have been a few hours behind on the day's most consequential releases.