NVIDIA Nemotron 3 Super is a 120B-parameter open hybrid MoE model, activating just 12B parameters for maximum compute efficiency and accuracy in complex multi-agent applications.
- Context window: 262,144 tokens
- Input: text
- Output: text
- Pricing: $0.08/M input tokens, $0.45/M output tokens
Community sentiment
Mixed (+3) — 2 positive, 2 negative, 9 neutral across 13 community samples from X and the web, as of 2026-10-02.
Feedback emphasizes Nemotron 3 Super’s strong reliability and usefulness as a teacher model (e.g., a reported 95.2% first-time usable-answer rate and success in distillation tests). Critics call out slow time-to-first-token (~19.3 s), missing coverage (e.g., IndicQA), and some RL experiments that expose reward-hacking behaviors.
- “When quality is measured only on successful calls, high scores can mask operational gaps. Route on first-time usable-answer rate alongside quality: NVIDIA Nemotron-3 Super 120B hits 95.2% of 2,466 attempts, delivering both reliability and performance.” — X
- “Groq (gpt-oss-20b)NVIDIA (nemotron-3-super)OpenRouter (gpt-4.1-mini) Live suite✅✅✅ Vendor error never reaches the tenant✅✅✅ Time to first token292 ms19.3 s1.1 s Reserved → billed (nano-USD)84,975 → 21,8250 → 0421,200 → 9,200 Prompt tokens, estimated + template → billed13 + 96 → 7” — X
- “Setup: three characters (pro-cheating / neutral / anti-cheating) distilled into Nemotron-3-Super, then RL on ImpossibleBench, where the only way to get the reward is to hack the tests. 3 runs each.” — X
- “(9) The same principle extends to other model families and other domains. We trained a Nemotron-3.5 student with a Nemotron-3-Super teacher and tested it on SWE-Bench Verified (real GitHub issues): Standard OPD: +0.2% PivotOPD: +3.2% (62.8 → 66.0), a third of the gap to the teach” — X
View on OpenRouter.
Model data sourced from OpenRouter.