--- title: "Perplexity Deploys NVIDIA's Nemotron Ultra to Run Agents 5x Faster" subtitle: "NVIDIA's new 550B Mixture-of-Experts model lands inside Perplexity's Pro and Max tiers, targeting multi-step agent workflows with 1M token context" url: "https://alphasignal.ai/news/perplexity-deploys-nvidia-s-nemotron-ultra-to-run-agents-5x-faster" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Perplexity" company_url: "https://alphasignal.ai/company/perplexity_ai" published: "2026-06-05T19:14:06.000Z" updated: "2026-06-05T19:43:26.466Z" type: "News" topics: - "Llms" - "Agents" subtopics: - "Long Context" - "Agent Frameworks" - "Computer Use" upvotes: 1032 read_time: "4 min" source_url: "https://x.com/perplexity_ai/status/2062976272436002825" comments: 0 --- # Perplexity Deploys NVIDIA's Nemotron Ultra to Run Agents 5x Faster > NVIDIA's new 550B Mixture-of-Experts model lands inside Perplexity's Pro and Max tiers, targeting multi-step agent workflows with 1M token context Perplexity has flipped the switch on [Nemotron 3 Ultra](https://developer.nvidia.com/blog/nvidia-nemotron-3-ultra-powers-faster-more-efficient-reasoning-for-long-running-agents/), NVIDIA's largest open model to date, for its Pro and Max subscribers in both the standard model picker and inside Perplexity Computer. The model is built around a very specific bet: that the next bottleneck in production AI is not single-turn chat quality but agents that run for hundreds of turns without losing the plot. ## A 550B model that only fires 55B at a time NVIDIA has released Nemotron 3 Ultra, a 550B total (55B active) open Mixture-of-Experts hybrid Mamba-Transformer for long-running agents. The architecture matters here. By interleaving Mamba layers, which scale linearly with sequence length, with traditional attention, NVIDIA gets a model that can hold a one-million-token architectural context window without the quadratic memory blowup a pure transformer would suffer. The reason for that design choice is the problem NVIDIA is targeting. Multi-agent workflows cause token counts to grow quickly. Agents plan, call tools, invoke sub-agents, receive information, and then pass history, outputs, and reasoning steps back into the model continuously. As tasks run longer, this constant communication increases costs and the risk of goal drift. Ultra is the orchestrator in that picture, with smaller Nemotron 3 models handling routine execution. ## The throughput pitch Speed is the headline number. Nemotron 3 Ultra is also fast. It achieves 5x higher throughput compared to other open models in its class, enabling long-running agents to complete tasks faster and more efficiently. NVIDIA also claims a meaningful cost delta: in experiments on SWE-bench and Terminal bench 2.0, it completed benchmarks using fewer total tokens and fewer tokens per turn than comparable models. This lowers the cost for agentic tasks by up to 30%. Some of that efficiency comes from quantization choices that are worth knowing about if you plan to self-host. NVIDIA ships a single NVFP4 checkpoint. On Blackwell it runs with native FP4 math. On Hopper it runs as W4A16, since Hopper lacks native FP4 tensor cores. The practical upshot: the reduced weight footprint has a deployment benefit. The W4A16 path leaves room to fit MTP weights on a single 8-GPU H100 node. An FP8 checkpoint could not, without spanning two nodes. ## Benchmarks, with caveats On coding-agent tasks, it scores 71.9 on SWE-Bench Verified and 56.4 on Terminal Bench 2.1. Long-context retrieval looks strong too: the model scores 94.7 on RULER at 1 million tokens. Several larger comparison models top out at 256K context. The harness-consistency story is unusual and worth pausing on. Most labs report a single peak score on a single agent scaffold. Nemotron Ultra is post-trained to deliver consistent accuracy across top harnesses. The result is SWEBench Verified scores between 65% and 70.4% across Pi, OpenHands, Hermes, OpenCode, and Mini SWE Agent, consistent performance regardless of which framework you deploy. Independent positioning is more sober. Ultra scores 48 on Artificial Analysis's Intelligence Index, ranking ninth of eighty-nine models and sitting well above the peer average of 31. Among US-origin open-weight models it leads the field, with the nearest peers being Gemma 4 31B at 39, the earlier Nemotron 3 Super at 36, and gpt-oss-120b at 33. Above it sits Kimi K2.6 at 54. ## What Perplexity users actually get Inside Perplexity, Ultra slots into the existing model selector alongside other frontier options. Perplexity has been steadily building out this approach: they post-train different open models for each stage of answering a question, from query parsing and retrieval to reranking and drafting the final response. This lets them precisely tune latency, cost, and relevance. The more interesting integration is Computer, Perplexity's agent product. It accepts a high-level goal, then automatically breaks it down into sub-agents, each working inside its own isolated cloud sandbox with access to a real filesystem and real browser. Ultra's harness-agnostic training and million-token context map directly onto that orchestration loop. ## If you want to skip Perplexity Distribution is wide on day one. Day-zero access spanned OpenRouter, NVIDIA NIM, Hugging Face weight downloads, Perplexity, Together AI, Fireworks AI, DeepInfra, Amazon SageMaker JumpStart, and more than twenty additional cloud and inference providers, over twenty-five platforms in total. The full package is released under a permissive license: a 550B total (55B active) open Mixture-of-Experts hybrid Mamba-Transformer for long-running agents that pairs a 1M-token context with up to ~6x higher inference throughput than comparable open LLMs at on-par accuracy, and ships with open weights, training data, and recipes under OpenMDW-1.1. For teams already paying for closed frontier APIs, the calculus has shifted slightly. You can fine-tune Ultra on proprietary data, run it behind your own firewall, and get harness-stable behavior across the agent frameworks most people are actually shipping. Whether the accuracy gap with closed models matters depends entirely on how many turns your agents need to survive before something breaks. ## Sources - [X post by @perplexity_ai](https://x.com/perplexity_ai/status/2062976272436002825) - [NVIDIA Nemotron 3 Ultra Powers Faster, More Efficient Reasoning for Long-Running Agents](https://developer.nvidia.com/blog/nvidia-nemotron-3-ultra-powers-faster-more-efficient-reasoning-for-long-running-agents/) - [NVIDIA Debuts Nemotron 3 Family of Open Models](https://nvidianews.nvidia.com/news/nvidia-debuts-nemotron-3-family-of-open-models) - [NVIDIA AI Releases Nemotron 3 Ultra: An Open 550B Mixture-of-Experts Hybrid Mamba-Transformer](https://www.marktechpost.com/2026/06/04/nvidia-ai-releases-nemotron-3-ultra-an-open-550b-mixture-of-experts-hybrid-mamba-transformer-for-long-running-agents/) - [Perplexity Joins the NVIDIA Nemotron Coalition](https://www.perplexity.ai/hub/blog/perplexity-joins-the-nvidia-nemotron-coalition) - [NVIDIA Nemotron 3 Ultra: 550B Open Reasoning Model Live](https://www.digitalapplied.com/blog/nvidia-nemotron-3-ultra-550b-open-reasoning-model-2026)