Nemotron 3 Super targets agentic reasoning

This is NVIDIA making a serious play for the “agent brain” layer: not just bigger reasoning, but cheaper, longer-context reasoning that can stay aligned across sprawling multi-agent workflows.

• –The 120B total / 12B active setup matters because it aims to keep inference costs down while still scaling to harder reasoning and coding tasks
• –A native 1M-token context window directly targets the context explosion problem that breaks many long-running agent systems
• –The hybrid Mamba-Transformer design, latent MoE, and multi-token prediction show NVIDIA optimizing for throughput as much as raw benchmark bragging rights
• –Open weights, datasets, and recipes make this more useful to developers than a closed API-only release, especially for teams that want to fine-tune or self-host
• –NVIDIA is also tying the model tightly to its own stack, from Blackwell NVFP4 optimization to NeMo, TensorRT-LLM, and NIM deployment paths