Nemotron 3 Super 120B A12B

Nemotron 3 Super 120B A12B is a reasoning and agentic model in NVIDIA's Nemotron 3 family that pairs a very large knowledge base with efficient sparse computation. Its hybrid LatentMoE design interleaves Mamba-2, mixture-of-experts, and attention layers and adds multi-token prediction on top, allowing 120 billion total parameters to be served while activating only about 12 billion per token. NVIDIA describes this combination as a way to deliver deep reasoning capacity without paying full dense-model compute cost on every forward pass, which makes the model well suited to long, multi-step tasks where most tokens are routine scaffolding and only selected steps require heavy reasoning.

For practitioners, the model offers configurable thinking that can be toggled on or off, a context window reaching one million tokens, and output lengths up to 32,768 tokens, which together support extended agentic workflows such as code synthesis across large repositories, multi-document analysis, and tool-driven problem solving. Training freshness is reported with a June 2025 pre-training cutoff and a March 2026 post-training cutoff, aligning the model's knowledge with recent developments. The architecture's focus on selective activation, long context, and explicit reasoning control makes it a strong fit for teams that need agent-style orchestration at scale rather than short conversational exchanges.