Model Drop

Nemotron 3 Super

March 11, 2026 2 min read

Nemotron 3 Super, which NVIDIA shipped on March 11, 2026, is an open-weight hybrid Mamba-Transformer mixture-of-experts model with 120B total and 12B active parameters and a 1M-token context, built for multi-agent workloads.

Model: Nemotron 3 Super (nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16)
Maker: NVIDIA
Ship date: March 11, 2026
Pricing: Free (open weights)
Context: 1M tokens
Where you can use it:
- Hugging Face in BF16, FP8 and NVFP4
- build.nvidia.com, Perplexity and OpenRouter
When to reach for it:
- Multi-agent apps with long, tool-heavy runs
- Skip it if top coding scores matter most
How it compares:
- Faster than gpt-oss-120b, per NVIDIA
- Trails Qwen3.5-122B-A10B on most evals
What makes it different:
- Mamba-Transformer hybrid with latent MoE
- Pretrained in 4-bit NVFP4, with MTP heads
Early reactions: Launch-day testers at Kilo called it a strong workhorse for scaffolding projects and summarizing docs, but found it weak with image-generation tools.
Read more: New NVIDIA Nemotron 3 Super Delivers 5x Higher Throughput for Agentic AI, Introducing Nemotron 3 Super, Model card, Technical report

The benchmarks

Scores are NVIDIA's BF16 checkpoint with reasoning on, against Qwen3.5-122B-A10B and gpt-oss-120b; NVIDIA used official rival scores where it could and filled the rest from aggregators or its own runs, and it ran SWE-bench Verified (OpenHands) and Terminal-Bench 2.0 on its own or official harnesses.

NVIDIA's numbers

SWE-bench Verified (real GitHub fixes, OpenHands): 60.47%. −5.93 under Qwen3.5-122B-A10B (66.40%), +18.57 over gpt-oss-120b (41.9%).
Terminal-Bench 2.0 (agentic terminal tasks): 31.00%. −6.50 under Qwen3.5-122B-A10B (37.50%), +12.30 over gpt-oss-120b (18.70%).
τ²-bench Telecom (customer-support tool use): 64.36%. −30.64 under Qwen3.5-122B-A10B (95.00%), −1.64 under gpt-oss-120b (66.00%).
RULER @ 1M (1M-token long context): 91.75%. +0.42 over Qwen3.5-122B-A10B (91.33%), +69.45 over gpt-oss-120b (22.30%).
HMMT Feb25 (competition math, no tools): 93.67%. +2.27 over Qwen3.5-122B-A10B (91.40%), +3.67 over gpt-oss-120b (90.00%).
GPQA Diamond (grad-level science, no tools): 79.23%. −7.37 under Qwen3.5-122B-A10B (86.60%), −0.87 under gpt-oss-120b (80.10%).
Humanity's Last Exam (no tools) (expert-level questions): 18.26%. −7.04 under Qwen3.5-122B-A10B (25.30%), +3.36 over gpt-oss-120b (14.90%).
IFBench (instruction following): 72.56%. −1.21 under Qwen3.5-122B-A10B (73.77%), +4.24 over gpt-oss-120b (68.32%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 12.8. +1.2 over gpt-oss-120b (11.6), −2.8 under Qwen3.5-122B-A10B (15.6).