As always, Nemotron 3 Ultra is fully open. This includes model weights, synthetic data, and post-training recipes. Available now on @huggingface →
AI
-
NVIDIA post-trains Ultra for popular agent harnesses
By
–
We post-trained Ultra for popular agent harnesses like @openclaw
, @NousResearch Hermes Agent, and @Langchain
. The result is an open frontier model developers can customize for specialized agents across domains. Read more: https://
nvda.ws/4adkn6J -
Ultra excels at large codebases, long reasoning chains, and multi-source synthesis
By
–
Beyond benchmark performance, Ultra can work through large codebases, reason across long chains of tool calls, and synthesize information gathered from hundreds of sources. pic.twitter.com/itDu34WVHk
— NVIDIA AI (@NVIDIAAI) 4 juin 2026Beyond benchmark performance, Ultra can work through large codebases, reason across long chains of tool calls, and synthesize information gathered from hundreds of sources.
-
Nvidia AI Ultra model excels at complex tasks with hybrid architecture
By
–
Ultra excels at complex tasks like coding and deep research. Long-running agents spend their time planning, using tools, recovering from failures, and deciding what to do next. The model’s hybrid Mamba-Transformer MoE architecture enables more reasoning cycles within the same
-
Nvidia Nemotron 3 Ultra: 550B MoE open frontier model
By
–
Today we're shipping Nemotron 3 Ultra.
— NVIDIA AI (@NVIDIAAI) 4 juin 2026
A 550B MoE frontier-intelligence open model built for long-running agents.
It delivers 5x faster inference and lowers the cost of complex agentic tasks by up to 30% versus other open frontier models. pic.twitter.com/FEXqvfzQFOToday we're shipping Nemotron 3 Ultra. A 550B MoE frontier-intelligence open model built for long-running agents. It delivers 5x faster inference and lowers the cost of complex agentic tasks by up to 30% versus other open frontier models.
-
DeepMind, IBM, Amazon admit using neurosymbolic; others don’t
By
–
DeepMind often acknowledges that they are using neurosymbolic in recent papers; IBM and Amazon, too. the others basically never no matter how much they use it
-
AI Performance as a System-Level Challenge: Optimizing Chips and Software
By
–
The enterprise takeaway is simple: AI performance is now a system-level challenge. The winners will optimize chips, memory, interconnects, software, and architecture together. Less latency means faster intelligence.
Less movement means lower cost.
Less waste means AI that can -
AI’s next bottleneck: time, not compute
By
–
AI’s next bottleneck is not just compute. It is time. Time lost moving data.
Time lost coordinating chips.
Time lost waiting on memory, interconnects, and software layers to catch up. That shift changes how we should think about AI infrastructure. A thread… #HuaweiPartner -
PlantOS: Correlating motor current, pressures, and descaling performance
By
–
PlantOS correlates motor current vs rolling force, hydraulic pressures vs strip dimensions, descaling performance vs surface drag patterns.
-
Expertise plus AI outperforms AI alone every time
By
–
This is why expertise plus AI beats AI alone every time. You know exactly what perfect looks like, so the model has a target. Most people don't have that, and it's obvious everytime.