Release of NVIDIA Nemotron 3 Ultra AI Model
NVIDIA released Nemotron 3 Ultra, a 550-billion-parameter open-weight mixture-of-experts (MoE) model designed for autonomous AI agents. While it leads other American open-source models, benchmarks indicate it remains behind Chinese competitors like Kimi K2.6.
Aggregated from 7 sources · Updated 25 Aug 2026, 03:07 UTC (UTC)
Coverage Balance
7 sourcesBlindspot Alert: Center Gap
This event is primarily covered by one side of the political spectrum. Niche or counter-narrative facts may be underrepresented.
Facts (10)
- Established
NVIDIA officially released Nemotron 3 Ultra, its most capable open-weight AI model, on June 4, 2026.
- Established
Nemotron 3 Ultra features 550 billion total parameters and utilizes a mixture-of-experts (MoE) architecture with 55 billion active parameters per token.
- Established
Nemotron 3 Ultra scores 47.7 on the Artificial Analysis Intelligence Index, outperforming other American open-weight models but trailing China's Kimi K2.6, which scores 53.9.
- EstablishedNiche Signal
Early enterprise adopters of Nemotron 3 Ultra include Perplexity, Palantir, ServiceNow, Glean, and CrowdStrike.
- Established
On August 11, 2026, NVIDIA released Nemotron 3.5 Lightning, a smaller 30B parameter model optimized for low-latency agent execution.
- Established
NVIDIA released the model weights, post-training recipes, and training corpora to Hugging Face and its own build.nvidia.com catalog.
- Emerging
The model supports a context window of up to 1 million tokens and utilizes a hybrid Mamba-Transformer architecture.
- Emerging
NVIDIA is collaborating with the Nemotron Coalition, including members like Nous Research and NAVER Cloud, to develop the foundation for a future Nemotron 4 family.
- EmergingNiche Signal
NVIDIA's cloud-compute budget for Nemotron development is reportedly capped at $7 billion through fiscal 2028.
- AttributionNiche Signal
NVIDIA claims the model delivers up to 5x faster inference and reduces costs for agentic tasks by up to 30% compared to previous iterations.