Nemotron 3 Ultra 550B (DeepInfra)

Tech Insider Canada reports that NVIDIA uploaded the Nemotron 3 Ultra weights to Hugging Face on June 4, 2026, following a Jensen Huang tease at the Computex keynote in Taipei. NVIDIA Research is cited as describing the model as a 550-billion-parameter mixture-of-experts with roughly 90% sparsity, activating about 55 b The article records an Artificial Analysis Intelligence Index score of 47.7 for Nemotron 3 Ultra, described as the top U.S. open-weight release but six points behind China's Kimi K2.6 at 53.9, with comparisons also drawn against Claude Opus 4.8 and Canadian sovereign-AI policy context (Ottawa's $2B AI investment). The

BenchmarkList publishes a third-party evaluation summary for NVIDIA's Nemotron 3 Ultra, aggregating 62 benchmark rows for the 550B mixture-of-experts release. The excerpt shows the model landing 7th of 73 on PinchBench Real-world OpenClaw with a best score of 90 (2026-06-04), 43rd of 326 on Artificial Analysis Terminal The page dates the benchmark records to early-to-mid June 2026 and frames the model as an NVIDIA open-weight flagship, but the excerpt does not explicitly tie any of these evaluation numbers or the listed token pricing to the DeepInfra-served variant or to LLM Gateway routing. Because the supplied evidence never confir