Nemotron 3 Super 120B-A12B
NVIDIA’s large hybrid mixture-of-experts model for reasoning and agent workflows. It needs server-class memory for the complete weights and a compatible runtime.
Exact local Ollama manifest and configuration checked on 2026-10-20. Memory figures are conservative planning estimates; no DeviceTerra inference or speed test is claimed. Requires Ollama 0.17.1 or newer. Hybrid Mamba2 MoE. Uses complete 120B weights despite 12B active. Registry context is 256K, below upstream 1M maximum. Store the complete 120B weights; 12B active does not mean a 12B download.
- Official model source
- Official engine artifact
Will it run?
MamiLens uses 118 GB as a conservative planning floor for this profile and its listed quantization. This is not measured speed. Processor, backend support, context length, memory bandwidth and other applications still affect the result.
Check your complete system →What is it useful for?
Suggested Ollama artifact
Confirm the tag on the official page before downloading, then run:
ollama run nemotron-3-super:120b-a12b-q4_K_MStart with a representative non-sensitive task. Record the exact runtime, quantization, context, load time, generation speed, memory use and answer quality.
Evidence and limits
Published context: 256K artifact maximum; practical context depends on memory. The source establishes specifications; it does not guarantee that maximum context fits at the MamiLens minimum RAM floor.