Inside Nvidia’s Nemotron 3 Super: A 120B-Parameter Hybrid Model Built for Multi‑Agent Workloads

Multi-agent AI systems are moving from research demos into production software engineering, cybersecurity, and operations workflows. But they come with a hard economic constraint: they can generate up to 15x the token volume of standard chat use cases, putting serious pressure on inference cost, latency, and infrastructure. Nvidia’s new Nemotron 3 Super model is explicitly aimed at this problem.

Nemotron 3 Super is a 120-billion-parameter hybrid model whose weights are available on Hugging Face under Nvidia’s custom open license. Architecturally, it combines state-space models, transformers, and a novel Latent Mixture-of-Experts (LatentMoE) to make long-context, agentic workloads more tractable on modern GPUs—especially on Nvidia’s Blackwell platform.

Why multi-agent workloads are breaking traditional LLM deployments

Enterprises experimenting with “agentic” architectures—chains of tools, planning agents, code-writing agents, and reviewers—are discovering a common bottleneck: context explosion. Each agent often pulls large documents, codebases, or logs into the prompt, and then passes expanded context to the next agent in the chain. The cumulative effect is a dramatic increase in tokens processed per task.

Compared with single-turn conversations, Nvidia reports that multi-agent systems can generate up to 15 times the token volume. For teams running on large proprietary models or high-parameter dense open models, this translates directly into higher GPU hours, higher cloud bills, and more complex capacity planning.

This is the backdrop for Nemotron 3 Super. Rather than building yet another dense transformer at 100B+ scale, Nvidia has designed a hybrid architecture that tries to preserve “big-model” reasoning ability while controlling memory usage and throughput under heavy, long-horizon workloads. The goal is not just strong benchmark scores, but practical throughput and context handling for multi-step research, code triage, and operations automation.

Inside the triple-hybrid: Mamba, Transformers, and LatentMoE

Nemotron 3 Super’s core architecture is a three-part hybrid: a Mamba-Transformer backbone plus a Latent Mixture-of-Experts layer design. Each component is aimed at a different pain point that shows up in multi-agent and long-context scenarios.

Hybrid Mamba-Transformer backbone. The model interleaves Mamba-2 state-space layers with strategically placed transformer attention layers. Mamba-2 provides linear-time sequence processing, which is important when context windows reach hundreds of thousands or millions of tokens. Nvidia positions Mamba-2 as a “fast-travel highway” that can process most of the sequence without the quadratic cost and growing KV cache associated with pure transformers.

However, pure state-space models typically struggle with associative recall—precisely recalling specific facts or lines of code from far back in the sequence. To address this, Nemotron 3 Super inserts transformer attention layers as “global anchors.” These layers help the model retrieve targeted information buried in long documents or large codebases, which is critical for tasks like reading a stack of financial reports or tracing a bug across a monorepo.

Latent Mixture-of-Experts (LatentMoE). Traditional Mixture-of-Experts architectures route tokens to subset “experts” at full hidden dimensionality, which can become a compute bottleneck as model size and number of experts grow. Nemotron 3 Super’s LatentMoE instead projects tokens into a compressed latent space before routing them to specialists.

This compression means the model can consult approximately four times as many experts for the same computational cost, giving it finer-grained specialization without blowing up FLOPs. Nvidia frames this as particularly important for agent use cases, where a single turn might require switching between Python generation, SQL reasoning, and natural-language explanation.

Multi-Token Prediction (MTP). On the decoding side, Nemotron 3 Super uses Multi-Token Prediction: instead of predicting just the next token, the model predicts several future tokens at once. Functionally, this acts as a built-in draft model for speculative decoding. Nvidia reports that this design enables up to 3x wall-clock speedups on structured generation tasks such as code or tool-call-heavy outputs—again targeting the multi-agent workload pattern where many calls are deterministic or semi-structured.

Performance profile: long context, reasoning, and throughput

Nvidia characterizes Nemotron 3 Super as a specialized agentic reasoning model rather than a general-purpose chat leader across every metric. On DeepResearch Bench—a benchmark for multi-step research over large document sets—the model currently holds the top position, aligning with its long-context design goals.

On a broader slate of benchmarks reported by Nvidia, Nemotron 3 Super lands in a competitive but mixed position relative to peer 120B-class models such as Qwen3.5-122B-A10B and GPT-OSS-120B. For example:

• On MMLU-Pro (general knowledge), Nemotron 3 Super scores 83.73, behind Qwen3.5’s 86.70 but ahead of GPT-OSS-120B’s 81.00.
• On math and reasoning tasks such as AIME25 (no tools) and HMMT Feb25, it is broadly in the same range as its peers, sometimes slightly ahead, sometimes just behind depending on the specific benchmark.
• On software engineering tasks, such as LiveCodeBench (81.19 vs Qwen3.5’s 78.93) and several SWE-Bench variants, Nemotron 3 Super scores below Qwen3.5 on some setups but substantially ahead of GPT-OSS-120B where comparable numbers exist.
• Long-context benchmarks such as RULER at 1M tokens show Nemotron 3 Super and Qwen3.5 performing similarly in the 90+ range, while GPT-OSS-120B falls significantly behind.
• In multilingual evaluation, Nemotron 3 Super’s MMLU-ProX score of 79.36 trails Qwen3.5’s 85.06 but beats GPT-OSS-120B’s 76.59.

For instruction following and chat, Nemotron 3 Super again tends to sit between Qwen3.5 and GPT-OSS-120B depending on the benchmark (e.g., 72.56 on IFBench prompt-based evaluation versus 73.77 for Qwen3.5 and 68.32 for GPT-OSS-120B).

Where Nemotron 3 Super clearly differentiates is throughput. In high-volume scenarios, Nvidia reports that it delivers up to 2.2x higher throughput than GPT-OSS-120B and 7.5x higher than Qwen3.5-122B. For organizations running fleets of multi-agent workflows, this speedup can be more impactful to total cost of ownership than a few percentage points on any single benchmark.

Blackwell-first design and NVFP4 efficiency

Nemotron 3 Super is engineered to take full advantage of Nvidia’s Blackwell GPU platform. The model was pre-trained natively in NVFP4, a 4-bit floating-point format optimized for Blackwell. According to Nvidia, this enables 4x faster inference on Blackwell compared with 8-bit models running on the prior Hopper architecture, without a loss in accuracy.

For enterprise teams, this matters on two fronts. First, NVFP4-native training reduces the need for post-hoc quantization experiments and mitigates some typical quality regressions associated with aggressive compression. Second, the reported 4x inference gain means that, on Blackwell, a 120B-parameter model can behave operationally more like a smaller dense model running on older hardware.

Since multi-agent workloads tend to be throughput-bound and latency-sensitive, especially when agents are orchestrated in parallel, this hardware-software co-design is central to Nemotron 3 Super’s value proposition. Organizations standardizing on Blackwell can potentially consolidate more agent runs per GPU and shrink the number of instances required to serve production workloads.

Licensing: “open weights” with guardrails and IP caveats

Nemotron 3 Super is released under the October 2025 version of the Nvidia Open Model License Agreement. Nvidia describes this as a permissive framework for commercial use, but it is distinct from OSI-style open-source licenses such as MIT or Apache 2.0.

Commercial usage and derivatives. The license explicitly states that models are “commercially usable” and grants users a perpetual, worldwide, royalty-free license to sell and distribute products built on top of Nemotron 3 Super. Enterprises may also create and own derivative models—fine-tuned variants or other modifications—provided they include the attribution: “Licensed by Nvidia Corporation under the Nvidia Open Model License.”

Output ownership. Nvidia disclaims ownership of outputs. Responsibility and ownership of generated content rest with the user, which is aligned with the expectations many enterprises now have around generative AI platforms.

Two key termination triggers. The license introduces two “red lines” that production teams need to track:

1. Safety guardrails. If a user bypasses or disables the model’s built-in “Guardrails” (safety mechanisms, limitations, or safety-related hyperparameters) without implementing a “substantially similar” replacement appropriate to the use case, the license automatically terminates.
2. Litigation trigger. If a user brings copyright or patent litigation against Nvidia claiming that the model infringes on their IP, their license to use the model terminates immediately.

In practice, this structure is designed to enable a broad commercial ecosystem while giving Nvidia legal and safety protections. For enterprise AI leaders, it means Nemotron 3 Super is not “open source” in the classic sense, but it is usable in closed-source products and internal systems as long as these conditions are respected.

Early ecosystem: agents, devtools, and industrial automation

Nvidia is positioning Nemotron 3 Super as part of its broader AI stack, rather than a standalone checkpoint. The model is available as an Nvidia NIM microservice, deployable on-premises through offerings like Dell AI Factory and HPE, and across major clouds including Google Cloud and Oracle, with AWS and Azure integrations listed as forthcoming.

On the application side, a set of early adopters reflects the model’s focus on complex, agentic workflows:

• CodeRabbit and Greptile are integrating Nemotron 3 Super for large-scale codebase analysis—scenarios where long context and fast traversal of repositories are critical.
• Siemens and Palantir are deploying the model to automate sophisticated workflows in manufacturing and cybersecurity, respectively, where multi-step reasoning and tool interaction are important.

Within the developer community, the release has attracted attention not only for the model itself but also for the transparency of its artifacts. Nvidia’s Chris Alexiuk highlighted the launch publicly, calling it a “SUPER DAY” and emphasizing that, in addition to model weights, Nvidia is releasing 10 trillion tokens of training data and recipes. He characterized Nemotron 3 Super as “FAST,” “SMART,” and “THE MOST OPEN MODEL WE’VE DONE YET,” underscoring Nvidia’s attempt to balance openness with the safeguards in its license.

What Nemotron 3 Super means for enterprise AI roadmaps

As enterprises move beyond chatbots into multi-agent applications, they encounter what Nvidia VP of AI Software Kari Briski calls “context explosion.” Nemotron 3 Super is Nvidia’s technical answer: a model that aims to deliver the “brainpower” of a 120B-parameter system with the effective operational profile of a smaller specialist, especially when paired with Blackwell GPUs.

For AI engineers and decision-makers, the trade-offs are relatively clear from the data Nvidia has published: Nemotron 3 Super is not the absolute benchmark leader on every axis, but it is highly competitive across reasoning, coding, and multilingual tasks while offering substantial throughput and long-context advantages over comparable open-weight models like GPT-OSS-120B and Qwen3.5-122B.

In concrete terms, organizations designing multi-agent systems, long-horizon research agents, or large-codebase copilots may find that the combination of 1M-token context support, hybrid Mamba-Transformer architecture, LatentMoE specialization, and NVFP4 efficiency on Blackwell changes the cost and latency envelope enough to make previously marginal use cases viable.

The licensing model, with its commercial permissiveness and safety/IP conditionals, will require legal review but is structured to enable closed-source, revenue-generating products. As multi-agent architectures continue to proliferate, Nemotron 3 Super’s balance of scale, throughput, and partial openness is likely to make it a prominent option in the open-weights segment of the enterprise LLM landscape.