Nemotron 3 Super

Nemotron 3 Super, released March 11, 2026, is NVIDIA's flagship language model that directly addresses two critical problems in production AI: the "thinking tax" (using massive reasoning models for every sub-task when most sub-tasks are simple) and "context explosion" (multi-agent systems generating 15x the tokens of standard chats, inflating costs and latency). With 120B total parameters but only 12B active per token, it achieves 2.2x higher throughput than GPT-OSS-120B and 5x throughput over the previous Nemotron Super.

The architecture introduces several innovations that work together for efficiency. Latent MoE routing compresses tokens before they reach the expert networks, allowing the system to call 4x as many specialist experts for the same inference cost. Multi-token prediction predicts multiple future tokens in a single forward pass, enabling built-in speculative decoding without requiring a separate draft model. The hybrid backbone interleaves Mamba-2 layers (which handle the majority of sequence processing with linear-time complexity) with Transformer attention layers at key depths for precise associative recall — combining the efficiency of state-space models with the precision of attention.

Native NVFP4 pretraining is a key differentiator. By training natively in 4-bit floating point on NVIDIA Blackwell GPUs, Nemotron 3 Super achieves a 4x speedup on B200 versus FP8 on H100 while maintaining accuracy. This is not post-training quantization — the model was pretrained in NVFP4 from the start, avoiding the quality degradation that typically accompanies aggressive quantization.

The model was RL post-trained using NeMo Gym and NeMo RL across 21 diverse environment configurations with over 1.2 million rollouts, giving it strong agentic reasoning capabilities grounded in real task completion rather than just instruction following. On PinchBench, it scores 85.6% — the best open model for the OpenClaw agent benchmark. It also leads on AIME 2025, SWE-Bench, and terminal-bench among open models.

The hybrid Mamba-Transformer backbone makes the 1M token context window practical. Mamba-2 layers handle long-range dependencies with linear-time complexity (O(n) rather than O(n²)), while Transformer attention layers interleaved at key depths provide the precise associative recall needed for tasks like needle-in-a-haystack retrieval and exact quotation. This hybrid approach delivers the context length of a Transformer with the throughput characteristics of a state-space model.

Nemotron 3 Super is particularly strong for agentic reasoning, software development, cybersecurity triaging, and multi-agent systems where throughput and efficiency matter. Its combination of hybrid architecture, native Blackwell optimization, massive RL training across 21 environments, and 5x throughput improvement over its predecessor makes it a unique offering in the open model ecosystem — delivering frontier-class agentic performance at a fraction of the inference cost of comparable dense models.