NVIDIA Launches Nemotron 3 Super: A 120B Open Model Built for Agentic AI
NVIDIA launched Nemotron 3 Super on March 11, 2026. The new open model combines 120B total parameters, 12B active parameters, a 1M-token context window, and higher inference throughput to target real production-grade agentic AI workloads.
NVIDIA officially introduced Nemotron 3 Super on March 11, 2026. In plain terms, this is an open model built for the kind of AI work that gets messy fast: long reasoning chains, tool use, multi-agent coordination, and code-heavy workflows that do not fit neatly into a short chat window.
On paper, the headline specs are straightforward: 120B total parameters, 12B active parameters, up to a 1M-token context window, and a stack that combines LatentMoE, multi-token prediction, NVFP4, and a hybrid Mamba-Transformer design. What makes the launch matter is the problem it is trying to solve. NVIDIA is not pitching another general-purpose chatbot. It is pitching infrastructure for agentic systems that need to stay fast once the context gets huge.
What NVIDIA Actually Released
Based on NVIDIA’s official blog, research page, model card, and technical report, Nemotron 3 Super comes with:
- 120B total parameters with 12B active parameters
- Up to a 1M-token context window
- A hybrid Mamba-Transformer architecture with LatentMoE
- Optimization for agentic reasoning, tool use, and RAG pipelines
- Open weights, plus supporting datasets, recipes, and technical documentation
That combination says a lot about where NVIDIA thinks demand is going. Teams are no longer only asking whether a model is smart enough. They are asking whether it can stay usable when an agent has to process a codebase, a long log trail, or a chain of tool calls without turning latency and cost into a problem.
If you want the broader system view, our earlier guides on building AI agents and MCP for agentic AI help explain the layer where a model like this fits.
Why This Model Is So Focused on Agentic AI
NVIDIA keeps returning to two phrases in its launch materials: context explosion and thinking tax. The first is what happens when agent workflows keep appending documents, tool traces, and intermediate steps until the prompt becomes massive. The second is the growing inference cost and delay that comes with deeper reasoning.
Nemotron 3 Super is clearly designed around those constraints. The 1M-token context window gives it room for long-horizon tasks. The 12B active-parameter setup is supposed to keep inference more efficient than the raw parameter count suggests. That matters for coding agents, enterprise search systems, security triage, and other workloads where one response is usually the least interesting part of the job.
The Performance Claims
This is where NVIDIA made its biggest swing.
According to the NVIDIA Research page and technical report, Nemotron 3 Super reaches:
- Up to 2.2x higher throughput than GPT-OSS-120B
- Up to 7.5x higher throughput than Qwen3.5-122B
Those figures come from an 8K input and 16K output benchmark setup. NVIDIA’s main launch blog frames the message a little differently, saying Nemotron 3 Super can deliver up to 5x higher throughput and up to 2x higher accuracy than the previous Nemotron Super in some scenarios.
Independent early coverage adds useful context. Artificial Analysis placed Nemotron 3 Super above GPT-OSS-120B on its own intelligence index, while still below a stronger Qwen 3.5 122B variant. That makes the picture more grounded. This is not obviously the strongest open model at everything, but it looks unusually compelling if efficiency and deployment economics matter as much as raw capability.
What Is Going on Under the Hood
The technical story is more than one trick. NVIDIA is stacking several choices that all point in the same direction.
NVFP4 Pretraining
Nemotron 3 Super is the first model in the family trained with NVFP4. That is also a hardware signal. NVIDIA says inference on Blackwell with NVFP4 can be up to 4x faster than FP8 inference on Hopper. So this launch is partly about the model, and partly about giving Blackwell-era deployment a strong software showcase.
LatentMoE
LatentMoE is meant to preserve the efficiency upside of mixture-of-experts designs while keeping long-context reasoning more practical. Buyers may not care about the term itself, but they will care if it shows up as lower inference cost and better response time in production.
Multi-Token Prediction and Hybrid Mamba-Attention
This matters for longer generations. Agent systems do more than emit short answers. They write plans, code, explanations, and tool traces. That makes output speed a bottleneck, so multi-token prediction and the hybrid sequence architecture are central to the pitch.
Early Adoption Signals
NVIDIA’s launch materials name Perplexity, CodeRabbit, Factory, Greptile, Palantir, Cadence, and Siemens among the companies already using or evaluating the model.
That list tells you what NVIDIA wants Nemotron 3 Super to be. This is not aimed first at consumer chat apps. It is aimed at the operational layer behind coding tools, enterprise software, engineering workflows, and industrial systems. Less flashy, maybe. Also much closer to paid deployment.
Why This Launch Matters
What stands out to me is not that NVIDIA released another open model. It is that the company bundled four priorities into one product story: open weights, ultra-long context, high throughput, and agentic workloads. That feels more practical than chasing a vague “best model overall” title.
A lot of models look great in demos. The harder question is whether they stay useful inside real agent systems, where prompts grow, tool calls multiply, and every extra second of latency becomes visible to the user and expensive for the team running it. Nemotron 3 Super is one of the clearer attempts to answer that problem directly.
The usual caveat still applies: official benchmarks are not the same as field performance. The real test will be reliability across external frameworks, tool-use accuracy, and how quickly errors stack up on long-running tasks. Even with that caution, Nemotron 3 Super already looks like one of the most important open-model releases to watch in March 2026.
Sources
- NVIDIA Blog: Introducing NVIDIA Nemotron 3 Super for Building Agentic AI Systems
- NVIDIA Developer Blog: Introducing Nemotron-3 Super, an Open Hybrid Mamba Transformer MoE for Agentic Reasoning
- NVIDIA Research: Nemotron-3 Super
- NVIDIA Build Model Card: Nemotron-3-Super-120B-A12B
- NVIDIA Technical Report PDF: NVIDIA Nemotron-3 Super
- Artificial Analysis: Nvidia Nemotron 3 Super, The New Leader in Open Efficient Intelligence?
Reader Dispatches
No reader responses yet. Share your technical perspective below.
Related Posts
AI News Roundup: The 5 Official Updates That Mattered Most in the Second Week of March 2026
As of March 15, 2026, the clearest AI trend is not just better models. This week's official updates from OpenAI, Google, Microsoft, NVIDIA, and Anthropic show the race shifting toward safety, multimodal retrieval, agent governance, and enterprise deployment.
Microsoft Pushes Copilot Into the Agent Era: What Wave 3, Agent 365, and the Frontier Suite Actually Mean
On March 9, 2026, Microsoft unveiled Copilot Wave 3, Agent 365, and the Frontier Suite. Here’s what changed, what it costs, and why the move matters for enterprise AI buyers watching governance, model choice, and lock-in.
Gemma 4 Is More Than an Open Model Release: Google Is Tying It to Local Agents and Android
Google released Gemma 4 on April 2, 2026. The bigger story is not just stronger open-model benchmarks, but Apache 2.0 licensing, a four-model split between large and edge tiers, and a direct push toward Android AICore and local agent workflows.