NVIDIA Nemotron‑3 Super 120B FP8: Massive Context & Agentic Reasoning Model

NVIDIA's Nemotron‑3 Super 120B‑A12B‑FP8 is a 120‑billion‑parameter large language model (with 12 B active parameters) released on March 11, 2026. Built on the Transformers library and tagged for text‑generation, it uses a hybrid LatentMixture‑of‑Experts (LatentMoE) architecture that interleaves Mamba‑2, MoE, and attention layers, and introduces Multi‑Token Prediction (MTP) for faster, higher‑quality generation. The model is quantized with NVIDIA's NVFP4 precision (FP8) to maximize efficiency on NVIDIA GPUs, requiring at least two H100‑80GB cards for inference.

The model supports up to 1 million tokens of context and seven languages (English, French, German, Italian, Japanese, Spanish, Chinese). Its design emphasizes agentic workflows: a configurable reasoning mode can be turned on or off via the chat template, enabling step‑by‑step thinking for tool use, long‑context reasoning, and RAG scenarios. Benchmarks show strong performance on a range of tasks, including MMLU‑Pro, HMMT, LiveCodeBench, and long‑context evaluations like RULER‑500, confirming its suitability for both general‑purpose chat and specialized high‑volume workloads such as IT ticket automation.

Nemotron‑3 Super is distributed under the NVIDIA Nemotron Open Model License, allowing commercial use. The README provides detailed deployment guides for vLLM, SGLang, and TRT‑LLM, along with a custom reasoning parser (super_v3) and examples of OpenAI‑compatible API usage. The model is linked to NVIDIA's post‑training and pre‑training dataset collections, and its training pipeline includes pre‑training, supervised fine‑tuning, and reinforcement learning with multi‑environment GRPO, all documented in the accompanying technical report.