Nemotron 3
NVIDIANVIDIA's first-party open-weight instruction-tuned model family for agentic reasoning, post-trained with SFT + RL on top of Nemotron Base. Current SKUs are Nemotron 3 Nano (30B total / 3B active hybrid MoE, March–April 2026) and Nemotron 3 Super (120B / 12B active Mamba-Transformer MoE, March 11, 2026); Nemotron 3 Ultra (550B total / 55B active) shipped 3 June 2026. Differentiates on (a) open training data and recipes published alongside the weights, (b) 1M-token context with strong RULER scores (Super 91.8 @ 1M), and (c) being NVIDIA's reference post-trained line for NeMo and NIM, optimized end-to-end against NVIDIA hardware. Also hosted free on OpenRouter and build.nvidia.com.
The Nemotron 3 family: Nano (30B total / 3B active hybrid MoE), Super (120B / 12B active hybrid Mamba-Transformer MoE, 1M context, 11 Mar 2026 at GTC) and Ultra (550B / 55B active, 3 Jun 2026). Designed for agentic/conversational use. Held a duplicate record in base_pretrained until 2026-07-29; the two were merged here because the family distributes post-trained weights only. The sole -Base repo on the Hub is nvidia/nemotron-3-8b-base-4k, created 2023-11-14 with a single download, which is the earlier 8B generation and not this family. Per the #114 convention, the observable artifact decides the category. BF16 originals are declared above; the NVFP4/FP8/GGUF quantizations are the same weights repackaged. An earlier note recorded the instruct repo as gated/401 -- all three resolve ungated. The score note used to call Super's SWE-bench 60.47 the highest open-weight figure; the report's own comparison column reads higher, and that claim has been withdrawn. Verified 2026-08-13 via the Nemotron 3 page, NVIDIA's developer blog and the Super technical report.
Openness
4 high confidence- weights
- open
- data
- open(~40M post-training/SFT+RL samples released as Nemotron-SFT-Data/Nemotron-RL-Data
- code
- open(training recipes + RL env configs)
- license
- NVIDIA-Nemotron-Open-Model-License(permissive,commercial,non-OSI)
Unusually open for a post-trained model: NVIDIA releases the weights, the post-training data and the recipes, but under the non-OSI NVIDIA Open Model License, so open weights rather than open source. This category scores the post-training data, which is genuinely released. The pretraining side is thinner than NVIDIA's "fully open" framing suggests - Nemotron-Pretraining-Specialized-v1.2 ships 53.6GB of synthetic data, but Nemotron-Pretraining-Code-v3 ships metadata parquet describing the code corpus rather than the corpus itself. That does not change the score, because the post-training answer is what this category asks for.
- https://research.nvidia.com/labs/nemotron/Nemotron-3/ recorded 2026-08-13
Nemotron 3 Nano/Super/Ultra family; the Data section still lists Nemotron-SFT-Data ("Collection of new Nemotron 3 Nano SFT datasets") and Nemotron-RL-Data alongside the pretraining corpora, followed by a Model Recipes section
- https://developer.nvidia.com/blog/introducing-nemotron-3-super-an-open-hybrid-mamba-transformer-moe-for-agentic-reasoning/ recorded 2026-08-13
NVIDIA primary blog - "Nemotron 3 Super is fully open-weights, datasets, and recipes"; license = NVIDIA Nemotron Open Model License (non-OSI)
- https://huggingface.co/api/datasets/nvidia/Nemotron-Pretraining-Specialized-v1.2 recorded 2026-08-13
92 files, 53.62GB, ungated. Nemotron-Pretraining-Fact-Seeking/*.parquet and related synthetic pretraining sets.
- https://huggingface.co/api/datasets/nvidia/Nemotron-Pretraining-Code-v3 recorded 2026-08-13
67 files, 1.20GB, ungated, all under Nemotron-Code-Metadata/*.parquet -- metadata describing the code corpus, not the corpus itself
Adoption
4 high confidenceAll three shipped repositories resolve ungated, and summing the family's distributed SKUs - which is what this category's adoption recipe asks for - the three BF16 repositories reach 2,226,055 downloads in the trailing 30 days (Nano 914,120; Super 870,782; Ultra 441,153), squarely in the 1M-10M band. The NVFP4, FP8 and GGUF quantizations are excluded to avoid counting the same weights twice. The family is also distributed free through OpenRouter (nemotron-3-super-120b-a12b:free) and NVIDIA build/NIM, which the download count does not capture, so this level is a floor.
- https://research.nvidia.com/labs/nemotron/Nemotron-3/ recorded 2026-08-13
official distribution via NVIDIA build/NIM + HF for agentic use
- https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 recorded 2026-08-13
914,120 downloads in the trailing 30 days, ungated
- https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 recorded 2026-08-13
870,782 downloads in the trailing 30 days, ungated
- https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 recorded 2026-08-13
441,153 downloads in the trailing 30 days, ungated
Capability
5 high confidenceFrontier-tier among open instruct models on the Super SKU. Nano is much smaller and would score lower, but the family's headline SKU defines the score. NVIDIA's own comparison table puts Qwen3.5-122B-A10B above Super on the same rows (SWE-Bench 66.40, MMLU-Pro 86.70), so Super is not the highest open-weight score on that comparison.
- https://research.nvidia.com/labs/nemotron/Nemotron-3/ recorded 2026-08-13
Nemotron 3 reasoning/agentic positioning - Nano "more accurate than GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507 on popular benchmarks"; Ultra released
- https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Super-Technical-Report.pdf recorded 2026-08-13
Agentic Reasoning Benchmark table, N-3-Super column - MMLU-Pro 83.73, AIME25 (no tools) 90.21, GPQA (no tools) 79.23, SWE-Bench (OpenHands) 60.47; the Qwen3.5-122B-A10B column on the same rows reads 86.70 and 66.40, and the GPT-OSS-120B column 81.00 and 41.9
Verified 2026-08-13