Nemotron 3 Super, NVIDIA’s latest open large language model, launched on March 11, 2026, targeting the growing demands of multi-agent artificial intelligence systems. The model carries 120 billion total parameters but activates only 12 billion during inference, delivering a balance between capacity and compute efficiency.
Multi-agent systems generate up to 15 times the tokens of standard chat sessions, re-sending history, tool outputs, and reasoning steps at every turn. This “context explosion” causes goal drift, where agents gradually lose alignment with the original objective. Additionally, deploying large reasoning models for every sub-task creates a “thinking tax” that makes multi-agent applications expensive and slow.
Architecture Behind Nemotron 3 Super
Nemotron 3 Super uses a hybrid Mamba-Transformer mixture-of-experts (MoE) backbone. Mamba-2 layers handle the majority of sequence processing using state space models, which provide linear-time complexity relative to sequence length. This design makes the model’s native 1 million-token context window practical for real workloads, such as reasoning over entire codebases or long document stacks.
Transformer attention layers are interleaved at key depths to preserve precise associative recall. MoE layers then scale the effective parameter count without the cost of dense computation, keeping latency low when many agents run concurrently.
Latent MoE and Multi-Token Prediction
The model introduces latent MoE, which compresses token embeddings into a low-rank latent space before routing decisions are made. This approach enables the model to consult four times as many expert specialists for the same inference cost as a standard MoE architecture. Finer-grained specialization becomes possible, with distinct experts activating for tasks such as Python syntax versus SQL logic.
Multi-token prediction (MTP) trains the model to forecast several future tokens simultaneously from each position. This produces measurable gains on chain-of-thought tasks and enables built-in speculative decoding at inference, delivering up to three times wall-clock speedups for structured generation tasks such as code and tool calls — without requiring a separate draft model.
Training Pipeline and Data Scale
Nemotron 3 Super is pretrained on 25 trillion tokens using NVFP4, NVIDIA’s 4-bit floating-point format optimized for NVIDIA Blackwell hardware. The pretraining corpus spans 10 trillion unique curated tokens. Training natively in reduced precision means the model learns accuracy within 4-bit arithmetic constraints from the first gradient update, rather than being quantized after training.
Supervised fine-tuning follows pretraining, using approximately 7 million samples drawn from a broader post-training corpus of 40 million samples covering reasoning, instruction following, coding, safety, and multi-step agent tasks. Reinforcement learning then refines the model across 21 environment configurations using NVIDIA NeMo Gym and NVIDIA NeMo RL, generating more than 1.2 million environment rollouts during training.
Benchmark Performance and Deployment
On PinchBench, a benchmark measuring how well language models perform as the reasoning core of an OpenClaw agent, Nemotron 3 Super scores 85.6% across the full test suite. NVIDIA states this makes it the top-performing open model in its class. The model also delivers over five times the throughput of the previous Nemotron Super, with four times improved memory and compute efficiency on NVIDIA B200 hardware compared to FP8 on NVIDIA H100.
“Nemotron 3 Super is ideal in this use. For instance, in software development, simple merge requests can be addressed by Nemotron 3 Nano while complex coding tasks that require deeper understanding of the code base can be handled by Nemotron 3 Super.”
Chris Alexiuk and Chintan Patel, NVIDIA
NVIDIA positions the model within a “Super + Nano” deployment pattern. Nemotron 3 Nano handles targeted, individual steps in an agentic workflow, while Nemotron 3 Super manages complex, multi-step planning and reasoning tasks. Expert-level computing tasks can be escalated further to proprietary models.
Open Weights, Datasets, and Recipes
Nemotron 3 Super is fully open, with model weights available on Hugging Face and through NVIDIA NIM. NVIDIA also releases the complete training and evaluation recipe, covering the full pipeline from pretraining through alignment. Deployment cookbooks are available for vLLM, SGLang, and NVIDIA TensorRT LLM, each with configuration templates and performance tuning guidance.
The open datasets include 10 trillion curated pretraining tokens, 40 million post-training samples, and approximately 10 of the 37 reinforcement learning datasets used during training. The model is also accessible through Baseten, Cloudflare, Coreweave, DeepInfra, Fireworks AI, Google Cloud, and several other cloud platforms. Fine-tuning cookbooks support LoRA, SFT, GRPO, and DAPO workflows using NVIDIA NeMo Megatron-Bridge and NeMo Automodel.