Multi-agent systems
Handle complex subtasks, long histories, and large tool libraries inside orchestrated agent workflows.
Independent tool overview
Nemotron 3 Super at a glance
NVIDIA Nemotron 3 Super is an open-weight 120B hybrid mixture-of-experts model built for long-context reasoning, tool use, coding, RAG, and agentic workloads.
Visit the official Nemotron 3 Super site ↗
- Developer
- NVIDIA
- Model size
- 120B total parameters; about 12B active
- Architecture
- Hybrid Mamba-2, LatentMoE, attention, and multi-token prediction
- Context window
- Up to 1 million tokens
- Primary inputs
- Text
- License
- NVIDIA Nemotron Open Model License
- Release date
- March 11, 2026
- Last reviewed
- August 30, 2026
Overview
What Nemotron 3 Super is
NVIDIA Nemotron 3 Super is a 120-billion-parameter language model designed for production agent systems rather than a consumer chatbot. Its sparse architecture activates about 12 billion parameters per token, combining Mamba-2, mixture-of-experts, attention, and multi-token prediction to reduce the cost of long reasoning and high-volume inference.
The model stands out for a native context window of up to one million tokens, configurable reasoning, tool calling, and openly released weights, data, and training recipes. Those advantages come with meaningful engineering requirements: the standard BF16 checkpoint needs substantial accelerator capacity, maximum context raises memory demands sharply, and teams remain responsible for evaluation, safeguards, orchestration, and operating costs.
Use cases
Who Nemotron 3 Super is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Repository-scale coding
Load large codebases or long technical context for code generation, debugging, and software-agent tasks.
Long-document RAG
Reason across extensive reports, knowledge bases, tickets, and retrieved evidence without aggressively fragmenting context.
High-volume automation
Use sparse activation and optimized precision to improve throughput for workloads such as IT ticket handling.
Private or customized deployment
Run open weights in controlled infrastructure and fine-tune or adapt the model for specialized domains.
Capabilities
Core Nemotron 3 Super features
120B sparse model
Nemotron 3 Super contains 120B total parameters while activating roughly 12B for each forward pass.
One-million-token context
The model supports up to 1M tokens for very large documents, repositories, histories, and retrieval sets.
Configurable reasoning
Developers can enable or disable thinking through the chat template to trade extra reasoning for speed and cost.
Tool-use training
The post-trained model is designed for function calling and agent workflows that interact with external systems.
Hybrid Mamba-Transformer design
Mamba-2 layers improve long-sequence efficiency while selected attention layers preserve global reasoning capacity.
Latent mixture of experts
A sparse expert design activates specialized capacity without paying the full dense-model inference cost per token.
Multi-token prediction
The architecture predicts multiple future tokens and includes a speculative-decoding head intended to increase generation throughput.
Multiple precision checkpoints
NVIDIA publishes BF16, FP8, and Blackwell-oriented NVFP4 variants for different deployment hardware.
Open deployment options
Weights can be served with supported frameworks or packaged through NVIDIA NIM for on-premises and cloud deployments.
Open training resources
NVIDIA publishes training data, post-training methods, reinforcement-learning environments, evaluation recipes, and technical reports alongside the model.
Seven post-training languages
The model card lists English, French, German, Italian, Japanese, Spanish, and Chinese as supported languages.
Process
How the Nemotron 3 Super workflow works
Step 1
Select a checkpoint
Choose BF16, FP8, or NVFP4 based on target hardware, accuracy requirements, and memory constraints.
Step 2
Review the license
Confirm the Nemotron Open Model License and any hosted-service terms fit the intended commercial or research use.
Step 3
Prototype through an API
Test prompts, reasoning modes, tool schemas, and long-context behavior with NVIDIA's trial endpoint or another supported provider.
Step 4
Build a representative evaluation
Measure task accuracy, tool-call validity, hallucinations, latency, throughput, and cost using real workload samples.
Step 5
Deploy the chosen serving stack
Run with a supported framework or NIM container, including NVIDIA's required reasoning parser and recommended generation settings.
Step 6
Add production controls
Wrap the model with access controls, input and output validation, observability, retries, human review, and safe tool permissions.
Step 7
Tune context and reasoning
Use the smallest useful context window and enable thinking only where quality gains justify the added latency and tokens.
Cost
Nemotron 3 Super pricing and free plan
The downloadable weights do not carry a per-token model fee, but every practical path has infrastructure or provider costs. NVIDIA does not publish one universal Nemotron 3 Super price because the model is available through self-hosting, an NVIDIA API trial, and multiple inference providers.
Open weights
No model download fee
Download and deploy under the NVIDIA Nemotron Open Model License.
- Infrastructure, storage, bandwidth, and engineering are paid separately
- The BF16 model card lists 8 H100 80GB GPUs as the minimum requirement
- Quantized variants can reduce hardware requirements
NVIDIA API trial
Trial access
Prototype against NVIDIA's hosted, OpenAI-compatible endpoint under API trial terms.
- Intended for evaluation and development
- Availability and quotas are shown in NVIDIA Build
- Production terms may require a separate NVIDIA service agreement
Hosted providers
Provider-specific
Use Nemotron 3 Super through participating clouds and inference platforms.
- Per-token or compute pricing varies by provider and precision
- Compare context limits, latency, data handling, and regional availability
- Verify the provider's exact model revision before production use
Pricing checked . Check current pricing at the source ↗
Assessment
Nemotron 3 Super strengths and limitations
Where it stands out
- One-million-token context supports unusually large working sets
- Only about one-tenth of the total parameters are active during inference
- Reasoning can be switched off for simpler, latency-sensitive tasks
- Purpose-built for tool use, coding, RAG, and multi-agent orchestration
- Open weights allow private deployment and customization
- BF16, FP8, and NVFP4 checkpoints cover multiple NVIDIA hardware generations
- Published datasets and recipes make the release more transparent and reproducible than weights alone
- OpenAI-compatible hosted access lowers the barrier to early evaluation
What to consider
- This is a developer model, not a polished end-user application
- The standard BF16 checkpoint requires expensive multi-GPU infrastructure
- Running the full 1M-token window sharply increases memory use, latency, and cost
- The default Hugging Face configuration uses 256K context rather than the maximum 1M
- Self-hosting requires expertise in inference serving, scaling, monitoring, and security
- Developers must use the required custom reasoning parser for supported backends
- Benchmark results do not guarantee performance on a team's real tasks or tool stack
- Reasoning traces and long outputs can add latency and token cost
- The model is text-focused rather than a native image, audio, or video model
- Seven primary post-training languages are narrower than some multilingual competitors
- Open weights do not remove the need for safety testing, access controls, and human oversight
- Hosted pricing, quotas, and available quantizations vary across providers
Compare
Nemotron 3 Super alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Coding
Meta Llama Models
Meta's Llama family offers a broader ecosystem of open-weight sizes and deployments for teams that do not need Nemotron's specific long-context agent focus.
Explore Meta Llama Models →Business Operations
DeepSeek
DeepSeek provides open reasoning models and hosted access that may be easier to evaluate for general coding and reasoning workloads.
Explore DeepSeek →Coding
Ollama
Ollama is a practical local model runner for teams exploring smaller open models before committing to Nemotron 3 Super's infrastructure footprint.
Explore Ollama →Questions
Nemotron 3 Super FAQs
What is Nemotron 3 Super?
Nemotron 3 Super is NVIDIA's open-weight 120B language model for agentic reasoning, tool use, coding, RAG, and high-volume automation. Its sparse architecture activates about 12B parameters per token.
Is Nemotron 3 Super open source?
NVIDIA describes the release as open and publishes weights, datasets, and recipes. Usage is governed by the NVIDIA Nemotron Open Model License, so it is more precise to call it an open-weight model and review that license for your use case.
How large is the context window?
The model supports up to one million tokens. The default Hugging Face configuration is 256K because longer contexts require substantially more memory.
Can Nemotron 3 Super run locally?
Yes, but not on an ordinary laptop in its standard form. NVIDIA lists eight H100 80GB GPUs for the BF16 checkpoint, while the NVFP4 variant can run on a single B200 or DGX Spark according to the official model card.
Does Nemotron 3 Super support tool calling?
Yes. It is post-trained for agentic workflows and tool use, but each integration still needs schema validation, permission controls, retries, and application-level safety checks.
Can I turn reasoning off?
Yes. The official chat template exposes an enable_thinking flag so developers can disable extended reasoning for simpler or more latency-sensitive requests.
How much does Nemotron 3 Super cost?
The weights have no per-token download fee under the model license. Actual cost depends on self-hosted GPU infrastructure or the pricing of the selected cloud or inference provider; NVIDIA Build also offers trial API access.
Which languages does Nemotron 3 Super support?
The model card lists English, French, German, Italian, Japanese, Spanish, and Chinese as supported post-training languages.
Is Nemotron 3 Super safe for autonomous agents?
No model should receive broad autonomous permissions solely because it scores well on tool-use benchmarks. Use least-privilege credentials, sandboxing, allowlisted tools, output validation, monitoring, and human approval for consequential actions.
Bottom line
Our Nemotron 3 Super verdict
Nemotron 3 Super is a compelling option for engineering teams that need open weights, very long context, and efficient reasoning inside sophisticated agent systems. It is best evaluated as infrastructure, not a chatbot: the model can lower inference cost relative to dense systems, but production value depends on hardware, serving expertise, task-specific evaluation, and strong controls around tools and data.
Visit Nemotron 3 Super website ↗Similar Consumer Tools
View all Consumer AI tools →
Copilot Cowork
Copilot Cowork - Microsoft's Anthropic-powered AI for running multi-step tasks across M365 apps
Mistral 4 Small
Mistral 4 Small - Mistral's open-source 119B MoE model combining reasoning, coding, and vision in one package
GPT 5.4
GPT 5.4 - OpenAI's new flagship reasoning model with native computer use and 1M context
GPT-5.4 mini & nano
GPT-5.4 mini & nano - OpenAI's fast, cheap small models built for coding agents and subagent workflows
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.