Nemotron 3 Super
Where NVIDIA's 120B open-weight agent model beats gpt-oss-120b, and where it trails it.
Nemotron 3 Super is a capable open-weight model for agentic execution and long-context processing at the 120-billion parameter scale. At Layer3Labs, we build custom agents for client workflow automation, where the model has to be fast enough at each step and still reason well enough to plan. The model uses 120 billion total parameters with 12 billion active per token, with up to 1 million tokens of context natively (most hosted endpoints serve 262K).
NVIDIA released the model to handle multi-step planning, whole-codebase software development, and large document analysis. Vendor benchmarks show strong throughput and long-context retrieval, though alternative open models lead on several standard evaluations.
Nemotron 3 Super Verdict
Nemotron 3 Super is a solid open-weight pick for agent and long-context work at its size. It is not a clear win over gpt-oss-120b, which beats it on several rows of NVIDIA's own benchmark table and costs far less to run per Artificial Analysis.
Three numbers carry the case. It has 120 billion total parameters with 12 billion active per token. It supports up to 1 million tokens of context natively, though most hosted endpoints serve 262K. NVIDIA reports up to 2.2 times the inference throughput of gpt-oss-120b at 8K input and 64K output tokens, but Artificial Analysis measured gpt-oss-120b faster on hosted endpoints (258.4 vs 140.3 tokens per second).
The architecture fits teams running high-throughput agent workflows, tool-calling pipelines, or codebase analysis on dedicated Hopper or Blackwell infrastructure. It is not suitable for teams requiring native vision processing, or teams that want the lowest cost per task on hosted endpoints. If you need image or video input, NVIDIA's newer Nemotron 3.5 Super VL adds vision and video. If hosted cost per task matters most, gpt-oss-120b is the cheaper pick on Artificial Analysis figures. We would change this verdict if a hosted price brought Nemotron 3 Super's $1.64 Artificial Analysis index run close to gpt-oss-120b's $0.09.
Architecture and Technical Specifications
Nemotron 3 Super uses a hybrid Latent Mixture-of-Experts (LatentMoE) architecture combining Mamba-2 state-space layers, Mixture-of-Experts (MoE) feed-forward layers, and standard attention layers.
NVIDIA announced the Nemotron 3 family on December 15, 2025, and published Nemotron 3 Super on March 11, 2026. The technical team submitted its technical report to arXiv on April 14, 2026.
The LatentMoE design projects tokens into a smaller latent dimension before expert routing. According to the NVIDIA developer blog, this lets the model run four times as many experts at the same inference cost. The architecture incorporates Multi-Token Prediction (MTP) layers with a shared-weight design across prediction heads. NVIDIA reports that MTP provides up to three times wall-clock speedups for structured outputs like code and tool calls without requiring an independent draft model.
Pre-training covered 25 trillion tokens drawn from a 10-trillion unique curated token corpus, with a data cutoff of June 2025. It represents the first model in the Nemotron 3 family pre-trained in NVFP4 for most linear layers, while latent projections, MTP layers, and attention projections remain in BF16 or MXFP8. Post-training data runs to February 2026. It used about 7 million Supervised Fine-Tuning (SFT) samples. Reinforcement Learning (RL) then ran about 1.2 million environment rollouts across 21 environment configurations in NeMo Gym and NeMo RL.
Weights are distributed under the NVIDIA Nemotron Open Model License, which permits commercial use and derivative works while retaining license notices. The NVIDIA NIM container ships under the NVIDIA Software License Agreement and Product-Specific Terms for AI Products.
Agentic Workflows and Workload Fit
NVIDIA designed Nemotron 3 Super for multi-step agent execution, structured tool invocation, Retrieval-Augmented Generation (RAG), and complex document reasoning.
Specific deployment targets include Information Technology (IT) ticket automation, multi-file software engineering, cybersecurity orchestration, and financial analysis over thousands of pages. The model incorporates reasoning controls via the enable_thinking toggle in its chat template, alongside configurable low_effort and reasoning_budget parameters. NVIDIA recommends temperature 1.0 and top_p 0.95. Supported languages are English, French, German, Italian, Japanese, Spanish and Chinese, plus programming languages.
In multi-agent systems, NVIDIA outlines a tiered deployment pattern pairing Nemotron 3 Super with smaller models. Under this pattern, Nemotron 3 Nano handles targeted steps, while Super handles complex planning and reasoning. If you want a smaller planner or a larger one, the Nemotron 3 overview covers Nano and the Nemotron 3 Ultra guide covers the top of the family.
Benchmark Performance and Evaluation Metrics
On NVIDIA's own table, Nemotron 3 Super's clearest lead is long context: 91.75 on RULER at 1M tokens against 22.30 for GPT-OSS-120B.
The Hugging Face model card publishes official vendor evaluations across core reasoning, agentic tool use, and codebase benchmarks. Testing was conducted using the NeMo Evaluator Software Development Kit (SDK) and NeMo Skills suite. The PinchBench row comes from NVIDIA's developer blog, not the model card. These are all NVIDIA's numbers. Independently, Artificial Analysis (read October 18, 2026) gives Nemotron 3 Super (Reasoning) an Intelligence Index of 13, rank #9 of 67 in its class, against 12 for gpt-oss-120b, with a 1.35-second time to first token.
Comparison Between Nemotron 3 Super and gpt-oss-120b
Nemotron 3 Super beats gpt-oss-120b on long context and agentic coding in NVIDIA's table, while gpt-oss-120b is faster and cheaper on hosted endpoints per Artificial Analysis.
Vendor benchmark tables published on the Hugging Face model card show a split outcome across standard evaluation suites. Nemotron 3 Super leads on SWE-Bench OpenHands (60.47 to 41.90), Terminal Bench Core 2.0 (31.00 to 18.70), and RULER at 1 million tokens (91.75 to 22.30). Conversely, gpt-oss-120b outperforms Nemotron 3 Super on LiveCodeBench v5 (88.00 to 81.19), Arena-Hard-V2 (90.26 to 73.88), and AIME25 without tools (92.50 to 90.21).
Inference performance also diverges depending on deployment topology. In NVIDIA's technical report, Nemotron 3 Super achieved up to 2.2 times higher inference throughput than gpt-oss-120b in a dedicated setting configured with 8,000 input tokens and 64,000 output tokens. On public hosted endpoints, however, Artificial Analysis recorded an output speed of 140.3 tokens per second for Nemotron 3 Super compared to 258.4 tokens per second for gpt-oss-120b.
Hosted pricing creates another distinction for organizations running heavy token volumes. To complete the Artificial Analysis Intelligence Index evaluation suite, Nemotron 3 Super required 170 million output tokens at a total cost of $1.64. By comparison, gpt-oss-120b required 84 million output tokens at a cost of $0.09. On the same Artificial Analysis pages, Nemotron 3 Super scores 13 on the Intelligence Index to gpt-oss-120b's 12.
Hardware Configurations and Local Serving
Local deployment of Nemotron 3 Super requires hardware matching the chosen precision format.
The 16-bit BF16 checkpoint requires a minimum of eight NVIDIA H100 80GB GPUs. It also fits across two Blackwell B200 or B300 GPUs using tensor parallelism. The FP8 checkpoint lowers the hardware floor to two H100 80GB GPUs or a single B200 GPU. The NVFP4 checkpoint runs on a single B200 GPU or a single DGX Spark. NVIDIA says NVFP4 on B200 runs up to four times faster than FP8 on H100.
- vLLM 0.18.1 serves the BF16 weights with
vllm serve --tensor-parallel-size 8 --max-model-len 262144 --enable-expert-parallel --reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder. Reaching the full 1 million context requires setting VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 and max-model-len 1048576. - SGLang supports serving via sglang serve with tp 8, ep 8, tool-call-parser qwen3_coder, and reasoning-parser nemotron_3.
- TensorRT-LLM version 1.3.0rc8 provides deployment using trtllm-serve with tp_size 8 and ep_size 8.
- Hugging Face Transformers supports the model natively starting in version 5.3.0. Earlier releases require setting trust_remote_code=True.
Hosted Endpoints and Cloud Pricing
Hosted endpoints provide Application Programming Interface (API) access to Nemotron 3 Super without the capital cost of dedicated GPU clusters.
As of October 2026, inference provider pricing varies based on precision and latency tiers. DeepInfra offers direct BF16 access at $0.085 per 1 million input tokens and $0.40 per 1 million output tokens on its Standard tier with a 262,144 token context limit. Its Priority tier costs $0.1275 input and $0.60 output, while Flex pricing is $0.068 input and $0.32 output.
OpenRouter routes to DeepInfra at $0.085 input and $0.40 output for BF16, DekaLLM at $0.08 input and $0.45 output for FP8, and Crusoe at $0.30 input and $2.40 output for FP8. Artificial Analysis lists $0.30 per 1 million input tokens and $0.90 per 1 million output tokens, with a blended price of $0.26. The Together AI and Baseten pricing pages list only Nemotron 3.5 Super VL, with no Nemotron 3 Super row.
The model initially shipped with availability on build.nvidia.com, Google Cloud Vertex AI, and Oracle Cloud, with deployment on AWS Bedrock and Microsoft Azure planned as follow-on releases.
Selection Tradeoffs and Newer Alternatives
Choosing Nemotron 3 Super depends on whether your workloads prioritize pure text-based reasoning over multimodal capabilities.
On October 15, 2026, NVIDIA released Nemotron 3.5 Super VL, expanding the architecture to include native vision and video processing. Artificial Analysis now marks Nemotron 3 Super as deprecated and names Nemotron 3.5 Super VL as its replacement. NVIDIA's own pages do not state a deprecation. Our detailed breakdown in Nemotron 3.5 Super VL vs Nemotron 3 Super compares the two models side by side.
For organizations running text-only agent pipelines on existing Hopper hardware, Nemotron 3 Super still has published BF16, FP8 and NVFP4 checkpoints and serving recipes for vLLM, SGLang and TensorRT-LLM. Audit your agent token usage, test prompt latency on DeepInfra or OpenRouter, and verify whether your workflow requires vision before provisioning dedicated hardware for Nemotron 3 Super.
What you need to run Nemotron 3 Super yourself
Nemotron 3 Super needs real memory, but it is within reach of a high-end workstation or a couple of professional GPUs — and many teams simply rent instead of buying. Match the path below to whether you want to own the hardware or pay by the hour.
| Path | What it is | Best for | Get started |
|---|---|---|---|
| Call the hosted API | Use Nemotron 3 Super as a pay-per-token API — zero hardware | Most teams; getting started | OpenRouter |
| Rent GPUs by the hour | Spin up an H100 / A100 for a few dollars an hour | Flexible self-hosting without buying cards | RunPod |
| Local on unified memory | One Mac with enough unified memory to hold a 4-bit quant | A single quiet on-prem box | Apple Mac Studio (M4 Max, 128GB) |
| Local on a workstation GPU | One 48GB pro card, or two 24GB consumer cards | Power users who want hardware they own | NVIDIA RTX 6000 Ada (48GB) |
To put Nemotron 3 Super to work once it is live, connect a coding client like Cursor (via OpenRouter) or a local runner such as Ollama.
Frequently Asked Questions
- Yes, for agent and long-context work at its size. It has 120B total and 12B active parameters and supports up to 1M tokens of context. NVIDIA reports 91.75 on RULER at 1M tokens and up to 2.2 times gpt-oss-120b's throughput. But gpt-oss-120b wins several rows of NVIDIA's own table and costs far less to run per Artificial Analysis, and NVIDIA has since released Nemotron 3.5 Super VL.
- Nemotron 3 Super is designed for complex agentic workflows, multi-file software engineering, Information Technology (IT) ticket automation, cybersecurity operations, and large-document Retrieval-Augmented Generation.
- Neither model is universally superior. Nemotron 3 Super outperforms gpt-oss-120b on SWE-Bench OpenHands (60.47 vs 41.90) and 1-million-token retrieval on RULER (91.75 vs 22.30), but gpt-oss-120b scores higher on LiveCodeBench v5 (88.00 vs 81.19) and cost $0.09 to run the Artificial Analysis Intelligence Index task set against $1.64 for Nemotron 3 Super.
- Nemotron 3 Super is an open-weight 120-billion parameter language model from NVIDIA that activates 12 billion parameters per token. It uses a hybrid Latent Mixture-of-Experts architecture combining Mamba-2, MoE layers, and Multi-Token Prediction.
- NVIDIA officially released Nemotron 3 Super on March 11, 2026, following the initial family preview announced on December 15, 2025.
- Video Random Access Memory (VRAM) requirements depend on precision. NVIDIA states minimum GPU configurations rather than a VRAM figure. The BF16 checkpoint needs at least eight 80GB H100 GPUs. The FP8 checkpoint needs at least two 80GB H100 GPUs or one B200. The NVFP4 checkpoint fits on a single B200 GPU or a DGX Spark.
Planning an Open-Model Rollout?
Layer3Labs helps teams pick, host and wire open models like Nemotron into real workflows, with the guardrails and cost checks done up front.
Book a Consultation