We’re excited to bring NVIDIA Nemotron 3 Super to Together AI, the AI Native Cloud. Built for multi-agent orchestration and complex reasoning, Nemotron 3 Super is a 120B-parameter (12B active) hybrid model that combines Transformer and Mamba architectures.
Running Nemotron 3 Super on Together AI Dedicated Inference allows engineering teams to deploy this open-weights model on managed infrastructure designed for high-throughput inference workloads.
Architectural capabilities for agentic workflows
Modern agentic systems that analyze massive document stores or orchestrate multi-step planning require models that can maintain state across long contexts without sacrificing generation speed. Nemotron 3 Super introduces several architectural innovations that make it well suited for these workloads:
- Hybrid MoE Architecture (Transformer + Mamba): By combining Mamba’s efficient sequence processing with Transformer attention, the model maintains strong reasoning capability while keeping active parameters (12B out of 120B) manageable for faster inference. Its Latent MoE design enables the model to call four experts for the inference cost of one, improving efficiency for reasoning-heavy workloads.
- 1M-Token Context Window: The 1-million-token context length allows applications to process entire codebases, maintain state across long agent trajectories, and inject significantly larger retrieval payloads directly into prompts.
- Multi-Token Prediction: Nemotron 3 Super is trained to generate several tokens simultaneously in a single forward pass. For applications that produce large outputs such as code generation or structured responses, this drastically reduces generation latency, delivering over 50% higher token generation speeds compared to current leading open models.
To achieve leading accuracy across benchmarks like AIME 2025 and SWE Bench verified, the model was trained using multi-environment reinforcement learning (RL) and NVIDIA-generated high-quality synthetic data. Because NVIDIA provides the model with open weights, datasets, and development recipes, engineering teams maintain full control to customize and fine-tune the model for their specific environments.
Running Nemotron 3 Super on Together AI
Serving a 120B-parameter hybrid model with a 1M-token context window typically requires distributed compute across multiple nodes. Nemotron 3 Super is available through Together AI Dedicated Inference, offering an infrastructure environment tailored for both experimentation and production scale without the overhead of GPU provisioning:
- Single-GPU Deployment: The model is optimized to run collaborating agents on a single GPU footprint, supporting deployment on single NVIDIA H200 or H100 GPUs. Together AI handles the underlying infrastructure orchestration, allowing teams to deploy these workloads without provisioning or managing GPUs directly.
- Research-Optimized Performance: Running hybrid MoE architectures efficiently requires highly tuned serving software. Together AI accelerates model execution through the Together Inference Engine and custom CUDA kernels. This stack helps teams achieve lower latency and higher throughput during live inference.
- Production-Grade Isolation: Dedicated Inference isolates workloads on reserved hardware to support predictable throughput and consistent performance at scale. The platform operates on enterprise-ready infrastructure, including a 99.9% uptime SLA and SOC 2 compliance.
Get Started
Developers can begin building with Nemotron 3 Super on Together AI today.
Run large-context reasoning workloads, deploy multi-agent systems, and scale production reference without managing GPU infrastructure.
FAQ
What is NVIDIA Nemotron 3 Super?
NVIDIA Nemotron 3 Super is a hybrid Mixture-of-Experts (MoE) reasoning model designed for complex AI workflows and multi-step problem solving. It combines Transformer and Mamba components to deliver strong reasoning capability with efficient inference.
What architecture does Nemotron 3 Super use?
Nemotron 3 Super uses a hybrid Mixture-of-Experts architecture that combines Transformer attention with Mamba sequence processing. This design improves compute efficiency while maintaining strong reasoning performance.
What context length does Nemotron 3 Super support?
Nemotron 3 Super supports context windows of up to 1 million tokens, enabling applications to analyze large document collections, maintain long conversations, and incorporate extensive retrieval context into reasoning workflows.
What types of applications can use Nemotron 3 Super?
Nemotron 3 Super is well suited for applications that coordinate multiple agents or operate across large knowledge sources. Examples include developer assistants that analyze and refactor codebases, enterprise systems that process large document collections, cybersecurity workflows that triage vulnerabilities or analyze system logs, and orchestration systems that route tasks across specialized agents based on user intent.
How do developers run Nemotron 3 Super on Together AI?
Nemotron 3 Super is deployed on Together AI through Dedicated Model Inference. Dedicated deployments allow teams to run models on reserved infrastructure designed for production workloads with predictable performance.
Do developers need to manage GPUs?
No. Together AI manages the underlying infrastructure, allowing developers to deploy and scale AI workloads without provisioning GPU resources directly.
Why use Together AI for these workloads?
Together AI provides infrastructure designed for large-scale AI systems, including reliable inference, serverless scaling, and managed infrastructure for modern AI applications.