Nemotron 3 Ultra 550B A55B

NVIDIA's Nemotron 3 Ultra is a 550B open-weight model released June 4, 2026, announced at Jensen Huang's Computex 2026 keynote, featuring 550 billion total parameters with only 55 billion active per token via Mixture-of-Experts at a 10:1 sparsity ratio. What distinguishes the architecture is the embedding of a Mamba-2 The article argues NVIDIA optimized Nemotron 3 Ultra for agents in production rather than for chatbot benchmarks, making it notable as an open-source model that addresses the economics of agentic systems making dozens of model calls per task. This is one of the most powerful open-source models available, with implicati

NVIDIA launched Nemotron 3 Ultra on June 4, 2026, at Computex, positioned as the top tier of a three-model Nemotron 3 family (Nano, Super, Ultra) designed for production agent workloads. The model is a hybrid Mamba-Transformer MoE with 550 billion total parameters and 55B active per token (a 10:1 sparsity ratio), train This practical walkthrough explains why the architecture matters for production agents rather than benchmark chasing, noting Ultra is intended as the orchestrator for hard calls: synthesizing contradictory evidence, sustaining architectural decisions across long sessions, and verifying complex designs against many cons

Nemotron 3 Ultra (550B A55B) holds the #1 rank on RULER with a score of 0.95/1 (long-context retrieval and multi-hop tracing benchmark), #1 on IMO-AnswerBench at 0.92/1 for mathematical reasoning on International Mathematical Olympiad problems, and #1 on PinchBench at 0.90/1 for coding agents on real-world software eng The aggregator positions Nemotron 3 Ultra at an LLM Stats Score of 36.3 with a blended price of $0.58 per million tokens, placing it between GPT OSS 120B and GPT-6 Astra on cost efficiency. The model's performance across conversation depths and capability tiers is tracked against frontier models on the leaderboard. Not

BenchmarkList reports third-party evaluation results for Nemotron 3 Ultra, including a score of 90 on PinchBench (ranking 7/73), 36.4% on Terminal-Bench Hard (ranking 43/326), 63.8 on MultiChallenge (ranking 7/34), and 83.3% on Tau2-Bench Telecom (ranking 80/332). Vals.ai Financial Agent 1.1 scored 60.1% without web se Evaluation dates span June 4-10, 2026, shortly after the model's June 4, 2026 release, providing early signal on agentic and tool-use performance across 62 reported benchmark rows. Listed pricing of $0.50 input / $2.20 output per 1M tokens reflects provider listings rather than creator-published rates.

Nemotron 3 Ultra requires substantial infrastructure for local deployment, with VRAM needs of approximately 319 GB at Q4_K_M, 390 GB at Q5_K_M, 588 GB at Q8_0, and 1100 GB at FP16, making it a server-class model even after aggressive quantization. CPU-only inference via llama.cpp or Ollama requires roughly 715 GB of sy Realistic throughput varies by hardware tier, with estimates of around 1.5 tokens/sec on entry-level GPUs, 2.5 tok/s on mid-range cards, and up to 5 tok/s on high-end hardware, assuming the chosen quantization fully fits in VRAM. The recommended local target is Q4 or Q5 quantization on multi-GPU H100/MI300 clusters, wi