Nemotron 3 Ultra

Nemotron 3 Ultra is NVIDIA's open-weight frontier reasoning model — a 550B total / 55B active parameter Mixture-of-Experts LLM built for agentic workflows including planning, tool use, coding, and long-horizon research tasks. It uses a hybrid Mamba-Transformer MoE architecture with LatentMoE and MTP layers for high-throughput inference on long contexts. Source: supabase.