Appendix A · NVIDIA Nemotron: Complete Open Roster
Research file for the model-making masterclass. Compiled November 17, 2026 (America/Chicago). Scope: every openly released Nemotron line relevant to open model building, from the 2023 Nemotron-3 8B through the Nemotron 3.5 and Nemotron Labs releases of late 2026.
How to read this file
- Facts were verified against primary sources where possible: NVIDIA developer and research pages, Hugging Face (HF) model cards, NGC catalog cards, arXiv tech reports, and NVIDIA GitHub repos.
- Where a fact rests on a secondary compilation (a catalog, wiki, or third-party analysis that quotes the card), the entry says so. Treat those as reported, not pinned.
- Benchmark numbers are vendor-reported unless a line says otherwise. The source is named for each headline figure.
- License names are per checkpoint. Nemotron does not have one family license. Six different NVIDIA license names appear across this roster, plus upstream licenses (Llama, Qwen, Gemma, Ministral) that stack on derivative checkpoints. Never infer a checkpoint’s license from the family name.
- Anything not confirmed at a primary source is collected again in the gaps section at the end.
License names that appear in this roster
- NVIDIA AI Foundation Models Community License Agreement: the 2023 Nemotron-3 8B line (NGC cards).
- NVIDIA Open Model License Agreement (also styled NVIDIA Open Model License): Nemotron-4 340B, the Llama-Nemotron derivatives, Nemotron 3 Nano and Super per their cards, Nemotron-Cascade-2 per secondary sources.
- NVIDIA Nemotron Open Model License: Nemotron 3 Nano 30B-A3B (HF card, last modified December 15, 2025), Nemotron 3 Super per card summaries, Nemotron-Labs-Sigma, Nemotron-Labs-Diffusion text models.
- OpenMDW License Agreement version 1.1 (OpenMDW-1.1): Nemotron 3 Ultra, Nemotron 3.5 Lightning, Nemotron 3.5 Super VL, both Nemotron 3.5 Bolt models, Nemotron 3.5 Transcribe, the Nemotron 3 Ultra teacher checkpoints, Nemotron-Parse-2.0.
- NVIDIA Internal Scientific Research and Development Model License: the license tag that HF metadata carries for at least the Nemotron-H-8B-Base-8K repo. This conflicts with downstream citations of the NVIDIA Open Model License for the same checkpoint. See the gaps section.
- NVIDIA Source Code License: the Nemotron-Labs-Diffusion 8B vision-language variant only, per the HF blog post.
- CC-BY-4.0: a few specialist checkpoints (OpenMath-Nemotron, OpenReasoning-Nemotron) and the HelpSteer2 dataset.
1. Nemotron-3 8B (2023): the first Nemotron
The name was reused in 2025 for an unrelated generation. The 2023 model is not an ancestor of the 2025 Nemotron 3 architecture.
- Release: November 15, 2023 (Wikipedia and NGC; the family launch blog is NVIDIA’s AI Foundation Models post).
- Checkpoints: Nemotron-3-8B-Base-4k, Chat variants (SFT, SteerLM, RLHF), and Nemotron-3-8B-QA-4k.
- Parameters: 8B, dense.
- Architecture: dense decoder-only Transformer (GPT-style), built for the NeMo framework.
- Context: 4,096 tokens.
- License: NVIDIA AI Foundation Models Community License Agreement (NGC cards for the QA and SteerLM checkpoints, read directly).
- Training material public: no standalone arXiv tech report was published for this checkpoint. NGC cards state the base was trained on about 3.5 trillion tokens (QA card); a secondary compilation says about 3.8 trillion tokens across 53 natural languages and 37 programming languages on 1,024 A100 GPUs for about 19 days. The token count discrepancy is in the gaps section.
- Headline claims: enterprise chat and QA positioning, SteerLM attribute steering at inference (quality, helpfulness, toxicity knobs). No benchmark table survives on the NGC cards read for this file.
- Status note: Microsoft Foundry lists the Chat-RLHF variant as retiring, with a pointer to migrate to Nemotron 3.5 Lightning.
2. Nemotron-4 (2024)
Nemotron-4-340B (Base, Instruct, Reward)
- Release: June 14, 2024 (NVIDIA research page and launch blog).
- Parameters: about 340B total (331.6B non-embedding), dense. No MoE.
- Architecture: dense Transformer.
- Context: 4,096 tokens.
- License: NVIDIA Open Model License Agreement (all three checkpoints).
- Training: 9 trillion tokens. Alignment data over 98 percent synthetic, generated with the Instruct and Reward models themselves.
- Public training material: tech report arXiv:2406.11704. HelpSteer2 dataset released under CC-BY-4.0. The release was positioned as a synthetic-data generation stack: the Reward model and the data pipeline were the product as much as the weights.
- Headline claims: the Reward checkpoint was reported number 1 on RewardBench at release. Two different figures (92.0 and 94.1) circulate in secondary sources; the discrepancy is in the gaps section.
Nemotron-4 15B
- A 15B base exists in NVIDIA’s research lineage: Nemotron-Mini-4B and Nemotron-4-Mini-Hindi-4B were compressed from a Nemotron-4 15B base. An open weight release of the 15B itself was not confirmed at a primary source in this pass. See gaps.
Minitron derivatives (2024)
Minitron is NVIDIA’s pruning plus distillation method. The models below are compressed derivatives of larger bases, released under the Minitron work.
- Llama-3.1-Minitron-4B (2024): pruned and distilled from Llama 3.1 8B. Dense, 128K context class.
- Mistral-NeMo-Minitron-8B-Instruct (October 2, 2024): compressed from Mistral NeMo 12B.
- Nemotron-Mini-4B and Nemotron-4-Mini-Hindi-4B (August to October 2024; Mini Hindi NIM October 23, 2024): compressed from the Nemotron-4 15B base, 4K context, NVIDIA Open Model License per secondary compilations.
- Public material: the Minitron approach is documented in NVIDIA’s pruning and distillation write-ups; per-checkpoint license names for the two Llama and Mistral derivatives stack NVIDIA terms on the upstream Llama 3.1 and Mistral terms. Exact card text was not re-read in this pass; see gaps.
3. Llama-Nemotron derivatives (late 2024 to 2025)
Two distinct waves: the 2024 RLHF wave on Llama 3.1, and the 2025 reasoning family built with neural architecture search (NAS) under the Puzzle framework.
Llama-3.1-Nemotron-70B-Instruct and 70B-Reward
- Release: October 15, 2024.
- Base: Meta Llama 3.1 70B Instruct. Post-training only; NVIDIA did not pretrain these.
- Parameters: 70B, dense. Context: 128K.
- Method: RLHF with REINFORCE on HelpSteer2 preferences.
- License: reported as NVIDIA Open Model License stacked on the Llama 3.1 Community License (secondary sources; the pattern is confirmed on the v1.5 Super card below).
- Public material: HelpSteer2 (CC-BY-4.0). The post-training approach fed directly into the Llama-Nemotron-Post-Training-Dataset release for the 2025 family.
- Headline claims (HF card, via secondary quotation): Arena Hard 85.0 against GPT-4o (May 2024) at 79.3; MT-Bench 8.98. Both were striking at release because the base was an open 70B.
Llama-3.1-Nemotron-51B-Instruct
- Release: September to October 2024.
- Base: derived from the 70B above by NAS plus block distillation. 51B, dense, 128K context class.
- Role: efficiency proof for the Puzzle NAS approach that the 2025 reasoning family then used at three sizes.
Llama-Nemotron reasoning family (2025)
Paper: “Llama-Nemotron: Efficient Reasoning Models”, arXiv:2505.00949 (May 2, 2025). Announced at GTC, March 2025. All three sizes are dense, Llama-derived, 128K context, with a system-prompt reasoning on and off toggle. Post-training: SFT, then Reward-aware Preference Optimization (RPO), then RL with verifiable rewards (RLVR, GRPO). The full post-training dataset is public as nvidia/Llama-Nemotron-Post-Training-Dataset on HF. NVIDIA’s Megatron Bridge documentation states that training recipes for Llama Nemotron are NOT currently available, so this family is open weights plus open post-training data, not open recipes.
License for the family: NVIDIA Open Model License Agreement, with the applicable Llama Community License stacked (confirmed verbatim on the Super v1.5 card: “GOVERNING TERMS: Your use of this model is governed by the NVIDIA Open Model License. Additional Information: Llama 3.3 Community License Agreement.”).
- Llama-3.1-Nemotron-Nano-8B-v1 (March 2025): from Llama 3.1 8B Instruct. 8B dense.
- Llama-3.1-Nemotron-Nano-4B-v1.1 (May 20, 2025): from Llama 3.1 Minitron 4B. 4B dense.
- Llama-3.3-Nemotron-Super-49B-v1 (March 2025): NAS derivative of Llama 3.3 70B Instruct, compressed to 49B so it fits a single H100 or H200 class GPU.
- Llama-3.3-Nemotron-Super-49B-v1.5 (July 25, 2025, HF): refresh of v1 with extended post-training (SFT, RPO, RLVR, iterative DPO for tool use). Card benchmarks (NeMo-Skills, reasoning on, averaged over up to 16 runs): MATH500 97.4, AIME 2024 87.5, AIME 2025 82.71, GPQA 71.97, LiveCodeBench (Oct 2024 to Feb 2025 window) 73.58, MMLU-Pro (CoT) 79.53.
- Llama-3.1-Nemotron-Ultra-253B-v1 (April 2025): NAS derivative of Llama 3.1 405B, compressed to 253B, sized to fit a single 8xH100 node. Paper benchmarks (reasoning on): GPQA-Diamond 76.0 (best open model on GPQA at the time, per the paper), AIME 2024 80.8, AIME 2025 72.5, MATH500 97.0. Note: a circulating secondary figure of AIME 2024 97.0 for this model conflicts with the paper table; the paper number is 80.8.
4. Nemotron-H (2025): dense hybrid Mamba-Transformer
Paper: arXiv:2504.03624 (April 2025). Nemotron-H is the dense hybrid line: Mamba-2 state-space layers do most of the work, only about 8 percent of layers are attention, and the family ships an FP8 training recipe. This is the architectural ancestor of the Nemotron 3 MoE hybrids.
- Nemotron-H-4B-Base-8K: 4B dense hybrid, 8K context. Listed in NVIDIA Megatron Bridge documentation.
- Nemotron-H-8B-Base-8K: 8B (8.1B per HF metadata), 52 layers with an interleaved Mamba and attention pattern (only 4 attention layers per the card description quoted by redistributors), 8K context. HF release around March 2025 per HF metadata crawls. A Nemotron-H-8B-Reasoning-128K variant also exists per secondary catalogs.
- Nemotron-H-47B-Base: pruned and distilled from the 56B with the MiniPuzzle method; secondary sources report about 20 percent faster inference than the 56B and about 63B distillation tokens.
- Nemotron-H-56B-Base: the full-size dense hybrid from the paper.
- License: CONFLICTED. HF metadata for Nemotron-H-8B-Base-8K carries the tag “nvidia-internal-scientific-research-and-development-model-license” (per a crawl of HF metadata), while downstream redistributors cite the checkpoint as “under the NVIDIA Open Model License”, and one secondary source calls the Nemotron-H cards “a research-focused NVIDIA license”. Do not treat these checkpoints as commercially cleared until the card text is read directly. See gaps.
- Public material: the paper discloses the architecture search and the FP8 recipe. NeMo and Megatron Bridge ship pretraining and finetuning recipes for the 4B, 8B, and 47B sizes (NVIDIA docs), which makes Nemotron-H one of the few lines with a vendor-supported from-scratch training path at small size.
5. Nemotron Nano v2 (2025): hybrid, FP8, built to be compressed
Paper: arXiv:2508.14444 (August 20, 2025).
- Nemotron-Nano-12B-v2-Base: 12B hybrid Mamba-Transformer, pretrained on 20 trillion tokens with an FP8 recipe.
- Nemotron-Nano-9B-v2: the aligned 12B compressed via Minitron to 9B. Exact parameter count 8,888,227,328. 56 layers: 27 Mamba-2, 25 MLP, 4 attention. 128K context. Sized to fit a single A10G (22 GiB) at BF16. A Nano-9B-v2-Base checkpoint was also released.
- License: reported as NVIDIA Open Model License in secondary sources; the exact card license string was not pinned in this pass. See gaps.
- Public material: the paper states the majority of the pre-training and post-training datasets were released on HF.
- Headline claims (paper and card, via secondary quotation of the card): AIME25 72.1, MATH500 97.8, GPQA 64.0, IFEval 90.3; up to 6x the throughput of Qwen3-8B in reasoning settings (paper claim).
6. Qwen-derived specialist reasoning lines (2025)
NVIDIA’s 2025 specialist program post-trained Qwen bases for single domains and published the datasets. These are post-training releases, not NVIDIA pretraining: the openness story is the data and the recipe write-ups.
OpenMath-Nemotron (7B, 14B, 32B)
- Release: April 23, 2025. Qwen2.5 derivatives tuned for competition math.
- License: CC-BY-4.0 per artifact listings (secondary).
- Public material: the OpenMathInstruct-2 style data pipeline and datasets are public on HF.
OpenCodeReasoning-Nemotron (7B, 14B, 32B)
- Release: April 25, 2025. Qwen derivatives tuned for code reasoning, distilled from DeepSeek-R1 traces.
- Public material: the OpenCodeReasoning dataset is public on HF.
AceReason-Nemotron (7B, 14B) and AceReason-Nemotron-1.1-7B
- Release: May 20, 2025 (1.1-7B followed June 16, 2025).
- Paper: arXiv:2505.16400 (NeurIPS 2025). Method: SFT from DeepSeek-R1-Distill-Qwen-7B and 14B, then GRPO reinforcement learning in two stages, math first, then code.
- License: NVIDIA Open Model License (secondary).
- Public material: dataset nvidia/AceReason-Math on HF, plus the paper’s full RL recipe description.
OpenReasoning-Nemotron (1.5B, 7B, 14B, 32B)
- Release: July 16, 2025. Qwen2.5-Instruct derivatives, SFT only (no RL), aimed at showing how far curated SFT data alone goes.
- License: CC-BY-4.0 (secondary).
- Public material: training data released as part of Nemotron-Post-Training-Dataset-v1; the science split is OpenScienceReasoning-2.
- Headline claims (card, via secondary quotation): the 32B posts 64.3, 73.1, and 80.0 across its headline reasoning table.
Nemotron-Research-Reasoning-Qwen-1.5B
- Release: May 29, 2025. A 1.5B research checkpoint from the same Qwen post-training program. Included for completeness; it is a research artifact, not a product tier.
7. Nemotron-Cascade (December 2025) and Cascade-2 (March 2026)
Cascade is NVIDIA’s post-training research line: the contribution is the training method (domain-wise Cascade RL, then multi-domain on-policy distillation), applied to other people’s bases.
Nemotron-Cascade-8B, 8B-Thinking, 14B-Thinking
- Release: HF December 8, 2025 (8B card date). Paper arXiv:2512.13607 (December 15, 2025).
- Bases: Qwen3-8B-Base and Qwen3-14B-Base. Dense. This is post-training only.
- Method: SFT, then domain-wise Cascade RL. The 14B-Thinking checkpoint is reported to beat its own SFT teacher, DeepSeek-R1-0528, on LiveCodeBench v5, v6, and Pro, and to have taken silver at IOI 2025 (paper claims).
- Public material: NVIDIA’s research page states the full models and the training data collection were released.
Nemotron-Cascade-2-30B-A3B
- Release: research introduction March 16, 2026; HF weights March 19, 2026. Paper arXiv:2603.19220.
- Base: the Nemotron 3 Nano base (30B total, 3B active, hybrid Mamba-Transformer MoE). NVIDIA’s own base this time, unlike the Qwen3-based first generation.
- Method: Cascade RL plus multi-domain on-policy distillation.
- License: reported as NVIDIA Open Model License (secondary; exact card string not pinned, see gaps).
- Public material: SFT and RL datasets public on HF.
- Headline claims (card, via secondary quotation): gold-medal level performance at IMO 2025 and IOI 2025; AIME 2025 92.4, rising to 98.6 with tool-integrated reasoning; LiveCodeBench v6 87.2.
8. Nemotron-Terminal (February 2026)
Paper: “On Data Engineering for Scaling LLM Terminal Capabilities”, arXiv:2602.21193 (February 24, 2026). A data-engineering release: the artifact is the dataset and the task-generation pipeline as much as the models.
- Nemotron-Terminal-8B, 14B, 32B: dense fine-tunes initialized from Qwen3 8B, 14B, and 32B, trained on Terminal-Corpus, a large synthetic terminal-task dataset produced by the Terminal-Task-Gen pipeline.
- License: per-checkpoint license not pinned in this pass (Qwen3 upstream terms are Apache 2.0, but the checkpoints’ own card terms were not read). See gaps.
- Public material: model checkpoints and most of the synthetic datasets are open-sourced in the HF collection nvidia/nemotron-terminal; the paper documents filtering, curriculum, and long-context training choices.
- Headline claims (paper, primary): Terminal-Bench 2.0 accuracy goes from 2.5 to 13.0 (8B), 4.0 to 20.2 (14B), and 3.4 to 27.4 (32B) over the Qwen3 bases. The 32B at 27.4 beats Qwen3-Coder-480B at 23.9 on the same bench.
9. Nemotron 3 (December 2025 to June 2026): the hybrid MoE generation
Family paper: “NVIDIA Nemotron 3: Efficient and Open Intelligence”, arXiv:2512.20856 (December 24, 2025). Announced December 15, 2025. Shared design: hybrid Mamba-2 and Transformer MoE, reasoning budget control, multi-environment RL post-training, context up to 1M tokens. Super and Ultra add LatentMoE and multi-token prediction (MTP) layers and were trained in NVFP4. NVIDIA’s stated family position: weights, datasets, and recipes all open.
Nemotron 3 Nano 30B-A3B
- Release: December 15, 2025.
- Parameters: 30B total (31,577,937,344 in the files per HF metadata), 3.5B active per the NGC card text. Some secondary sources compute 3.2B active excluding embeddings; the card’s own headline is 3.5B.
- Architecture: hybrid MoE, 52 layers total: 23 Mamba-2 and MoE layers plus 6 attention layers; 128 routed experts plus 1 shared expert per MoE layer, top-6 per token. Note: the NGC card text is internally inconsistent on routing (one line says 5 experts activated per token, another says top-6). See gaps.
- Context: up to 1M tokens per the card; the shipped config defaults to 262,144 positions.
- License: NVIDIA Nemotron Open Model License (HF card, last modified December 15, 2025). The NGC card for the same checkpoint says “NVIDIA Open Model License Agreement”. Both names are NVIDIA’s own; the HF card name governs the HF download.
- Training: 25 trillion tokens (card) drawn from a released pretraining corpus whose disclosed total is 10,648,823,153,919 tokens (about 10.6T) in the Nemotron-Pre-Training-Datasets collection, so the run is multi-epoch over the released corpus. Post-training: 13M+ SFT samples (card, via secondary quotation), RL via NeMo Gym, with Qwen3-Nemotron-235B-A22B-GenRM as the reward model.
- Public material: pretraining dataset collection, post-training datasets, tech report, and recipes via the NVIDIA-NeMo/Nemotron GitHub repo.
- Headline claims (card, via secondary quotation): RULER 92.9 at 256K context and 86.3 at 1M.
Nemotron 3 Nano 4B
- Release: March 16, 2026.
- Parameters: 3.97B, dense hybrid (Mamba-2 plus MLP, 42 layers, 4 attention layers). Not an MoE.
- Method: compressed from Nemotron-Nano-9B-v2 with the Nemotron Elastic framework.
- Context: 262,144.
- Targets: edge hardware (Jetson Thor, RTX, DGX Spark).
- License: NVIDIA Nemotron Open Model License (card, via secondary system-card mirrors).
- Headline claims (card, via secondary quotation): BFCL v3 61.1, IFEval 82.8.
Nemotron 3 Nano Omni 30B-A3B
- Release: April 28, 2026. Paper arXiv:2604.24954 (per HF tags, via secondary).
- What it is: the Nano 30B-A3B architecture extended to multimodal reasoning over video, audio, image, and text.
- License: HF tag reported as “nvidia-open-model-agreement” (secondary). Exact card terms not pinned; see gaps.
Nemotron 3 Super 120B-A12B
- Release: March 11, 2026. Tech report: NVIDIA research PDF dated April 3, 2026, also on arXiv as 2604.12374.
- Parameters: 120B total, 12B active nominal (12.7B including embeddings per one analysis of the report).
- Architecture: hybrid Mamba-2 and attention LatentMoE with MTP layers. First Nemotron 3 model pretrained end to end in NVFP4 (most linear layers NVFP4; latent projections, MTP, and attention projections in BF16 or MXFP8).
- Context: up to 1M tokens.
- License: NVIDIA Nemotron Open Model License (card, via secondary summaries; direct card read still owed, see gaps).
- Training: 25 trillion tokens in two phases (20T diversity, 5T high quality) from about 10T unique curated tokens. Post-training: SFT on about 7M samples, then multi-environment RLVR across 21 environment configurations with about 1.2M rollouts (NeMo Gym and NeMo RL), then SWE-RL, then RLHF, then MTP healing.
- Public material: the tech report states the datasets, base, post-trained, and quantized checkpoints are open-sourced on HF; the post-training data ships as the Nemotron-Super-Post-Training-Data collection (RL environments plus SFT datasets); the training recipe is shared through the Nemotron GitHub repo.
- Headline claims (tech report): SWE-Bench Verified 60.47 (OpenHands harness), AIME 2025 89.0, GPQA (no tools) 79.23, MMLU-Pro 83.73, RULER at 1M 91.75; throughput up to 2.2x GPT-OSS-120B and 7.5x Qwen3.5-122B at 1M context.
Nemotron 3 Ultra 550B-A55B
- Release: June 4, 2026. Repos: BF16, NVFP4, Base-BF16, and a GenRM (generative reward model) checkpoint.
- Parameters: 550B total, 55B active.
- Architecture: hybrid Mamba-2, MoE (LatentMoE), and attention, with MTP.
- Context: 262K in BF16; 1M in NVFP4 on Blackwell.
- License: OpenMDW License Agreement version 1.1 (OpenMDW-1.1), confirmed on the HF card and the build.nvidia.com card.
- Training: the HF card records about 20 trillion pretraining tokens in NVFP4. Post-training includes MOPD (multi-teacher on-policy distillation) from 10+ domain-specialized teacher checkpoints built on the Ultra recipe. The Base card documents the released corpus (Nemotron-CC and Common Crawl snapshots) plus 10M SFT samples and 1M RL tasks published under OpenMDW-1.1 (per a secondary summary of the Base card).
- Headline claims (HF BF16 card): SWE-Bench Verified 70.7, SWE-Bench Multilingual 67.7, GPQA (no tools) 87.0, LiveCodeBench v6 89.0, Humanity’s Last Exam 26.7. The build.nvidia.com card gives SWE-Bench Verified 71.9 for BF16 and 69.7 for NVFP4, a small cross-surface discrepancy worth knowing. Artificial Analysis scored it 48 on its Intelligence Index, reported as the top US-origin open-weight model at the time.
- Note on the pretraining token figure: a developer-blog-derived secondary summary describes a 10T base plus 212B new tokens, which conflicts with the card’s about 20T. The card figure is the one to cite; the conflict is logged in gaps.
NVIDIA-Nemotron-Labs-Teacher-Competition-Coding (550B-A55B)
- Release: HF August 14, 2026 (card read via search result).
- What it is: one of the 10+ domain-specialized teacher checkpoints from the Nemotron 3 Ultra recipe, released openly. 550B total, 55B active, LatentMoE Mamba-2 plus MoE plus attention hybrid with MTP. Model dates December 2025 to May 2026.
- License: OpenMDW-1.1.
- Why it matters for builders: the MOPD teacher set is itself partially public, which is rare. A student model can be distilled from the same teachers NVIDIA used.
10. Nemotron 3.5 (August to October 2026)
The 3.5 step keeps the hybrid MoE architecture and moves the whole line to OpenMDW-1.1. A technical report titled “NVIDIA Nemotron 3.5 Super VL and Lightning Technical Report” covers the two leads.
Nemotron 3.5 Lightning 30B-A3B
- Release: August 11, 2026.
- Parameters: 30B total, about 3B active. Architecture tag nemotron_h: interleaved Mamba-2, MoE, and selective attention layers with one MTP layer; 128 routed experts, 6 active per token (NeMo docs).
- Context: up to 1M tokens in NVFP4 on Blackwell; 256K on a single H100 in BF16.
- License: OpenMDW-1.1 (multiple secondary sources citing the card).
- Ships as: BF16, NVFP4, NVFP4-DSpark and DFlash speculative-decoding variants, plus a Base-BF16 checkpoint. Launched alongside NeMo Switchyard, an open-source router.
- Positioning: agent-harness execution steps and a post-training base. Card benchmark table (via secondary quotation) includes SWE-bench Verified 51.56, against 70.12 for Qwen3.6-35B-A3B in the same table; the card sells throughput and cost per task, not peak accuracy.
- Reported downstream use (held as reported, not verified here): CodeRabbit with NVIDIA and Baseten post-trained Lightning for code-review routing (SFT on 9,996 examples plus GRPO, 80.7 percent routing accuracy, under 100 dollars of training cost).
Nemotron 3.5 Super VL 120B-A12B
- Release: HF October 15, 2026. Card read directly for this file.
- Parameters: 120B total, 12B active. 512 routed experts per MoE layer.
- Architecture: Mamba-2 and Transformer hybrid MoE plus a C-RADIO vision encoder; MTP for native speculative decoding. Inputs: text, image, video.
- Context: up to 1M tokens.
- License: OpenMDW-1.1 (card’s governing terms link).
- Training (card): backbone initialized from Nemotron 3 Super Base; MTP head trained on 100B tokens; then 1T tokens of distillation from Nemotron 3 Ultra Base; 100B tokens of Indic continued pretraining (the synthetic Indic corpus of about 2.83T tokens in 22 languages, co-created with IIT Madras, was released with this model); a 35B-token extension to 1M context; 67B tokens of software-engineering mid-training. Card pretraining corpus table: Nemotron-CC-v2 and v2.1 9.1T, Nemotron-CC-Code-v1 427.9B, Nemotron-Pretraining-Code v1 to v3 1.7T, Nemotron-CC-Math-v1 133.3B, Nemotron-Pretraining-Specialized plus SFT sets 660B.
- Headline claims (card, BF16): GPQA-Diamond 89.52, LiveCodeBench v6 88.22, SWE-bench Verified 71.10, BrowseComp 78.
Nemotron 3.5 Bolt 74B-A9B (24GB)
- Release: October 15, 2026, alongside Super VL.
- Parameters: 74B total, 9B active. Multimodal (text, image, video up to 2 minutes). Context: 256K.
- Method: Super VL compressed by architecture search (routing cut to 9 experts), retrained on 120B tokens, then RL for tool use. The shipping artifact stores experts at about 2 bits (IQ2_XXS, 2.0625 bits per weight), so the GGUF is about 22 GB and the pitch is a 24 GB memory budget on RTX Spark, DGX Spark, and GeForce RTX. Caveat from independent coverage: NVIDIA’s own deployment table asks for a 32 GB GPU at 65K context and 48 GB for the full 262K context, so 24 GB is a weights budget, not a working configuration.
- License: OpenMDW-1.1.
- Headline claims (card, via secondary quotation): Terminal-Bench 2.1 50.98 against Qwen3.6-35B-A3B 44.38 and Gemma 4 26B-A4B 37.22 (both BF16, uncompressed); BrowseComp 66.75 against 48.74 and 26.30; PinchBench 87.82, just behind Qwen’s 88.07 in the same table. Artificial Analysis Intelligence Index 21.4, reported fourth among models under 100B.
Nemotron 3.5 Bolt E4B (4GB)
- Release: October 23, 2026.
- Parameters: 8B total, 4B effective.
- Base: NOT a Nemotron pretrain. NVIDIA post-trained Google’s Gemma 4 E4B-it with RL for multi-turn tool use, instruction following, and short spoken replies. Text only, 64K context in and out. Ships as a Q4_K_M build of about 4 to 6 GB for RTX, Jetson, and Spark class devices.
- License: OpenMDW-1.1 for NVIDIA’s terms, stacked on Gemma 4’s Apache 2.0 upstream terms.
- Builder note: the E4B name hides the lineage. Anyone evaluating it as evidence of NVIDIA’s small-model pretraining is making a category error; it is a post-training release on a Google base.
Nemotron 3.5 Transcribe 30B-A4B
- Release: October 28, 2026. Card read via search result for this file.
- Parameters: 34B total including the audio encoder, about 4.3B active. Artificial Analysis counts it as 30B total and 4B active, following the repo name; the card’s 34B figure includes the encoder stack. Both numbers are in circulation; the card is authoritative.
- Architecture: Mamba-2 and Transformer hybrid MoE language backbone (the Nemotron 3.5 Lightning decoder) plus a Parallel Expert Encoder: a trainable Transformer ASR encoder (about 600M) and a frozen Sortformer speaker encoder (about 100M). Offline only, recordings up to 1 hour, 65,536-token context, 40 languages.
- License: OpenMDW-1.1.
- Public material: trained on Granary-v2, an open NVIDIA speech dataset (843,600 audio hours across 48 languages; the public manifest covers about 463,062 hours). A technical report accompanies the release.
- Headline claims (card, results captured October 12, 2026, BF16): English ASR mean WER 4.71 across 9 Open ASR Leaderboard sets (next best open weight 4.94); multilingual mean WER 4.62 across 7 languages; long-form mean WER 10.98; multi-speaker mean cpWER 15.80. Ships in BF16, FP8, and NVFP4.
Smaller 3.5 side lines (noted, lightly verified)
- Nemotron 3.5 Content Safety (March 16, 2026): 4.3B guard model per a secondary catalog.
- Nemotron 3.5 Embed 1B and 8B: embedding models listed on OpenRouter’s NVIDIA page (secondary listing only).
- Nemotron 3.5 VoiceChat: named in a secondary compilation as OpenMDW-1.1. No card was read for this file.
- Nemotron-Parse-2.0 (June 30, 2026): document parsing model, OpenMDW-1.1 per HF listing data (secondary).
11. Nemotron Labs (2026): the diffusion research lines
Labs is NVIDIA’s research label. Two distinct diffusion families sit under it, and they are easy to confuse. Nemotron-Labs-Diffusion (NLD) is a masked-diffusion, tri-mode family built on Ministral 3. Nemotron-Labs-Sigma is a continuous-diffusion family, warm-started from autoregressive weights, released quietly with no announcement.
Nemotron-Labs-Diffusion (NLD) 3B, 8B, 14B, plus 8B VLM
- Release: research page May 19, 2026; HF blog May 23, 2026; arXiv preprint followed in July 2026. HF collection: nvidia/nemotron-labs-diffusion.
- Architecture: dense Transformers initialized from Mistral Ministral 3 checkpoints, trained first on a purely autoregressive objective (about 1T tokens), then on a joint AR plus block-diffusion objective (about 300B tokens). Total pretraining 1.3T tokens from the NVIDIA Nemotron Pretraining datasets (HF blog), then SFT. One checkpoint serves three decoding modes: autoregressive, block diffusion (32-token blocks), and self-speculation (diffusion drafts, AR verifies). An optional LoRA drafter adapter (about 36M parameters on the 8B) boosts tokens per forward.
- Variants: base and instruct at 3B, 8B, and 14B, plus an 8B vision-language model built from Ministral3-8B-Instruct.
- License: text models under the NVIDIA Nemotron Open Model License (HF blog); the 8B VLM under the NVIDIA Source Code License (HF blog). Checkpoints require trust_remote_code.
- Public material: training recipe and code on GitHub (linked from the HF blog); tech report PDF on NVIDIA’s site.
- Headline claims (NVIDIA research page and HF blog): the 8B decodes about 6x more tokens per forward than Qwen3-8B with slightly better average accuracy (plus 1.2 points across its 10-task instruct eval), translating to about 4x throughput on SPEED-Bench with SGLang on a GB200. On a DGX Spark, the 8B card reports 112 tokens per second in self-speculation mode against 41.8 autoregressive (via secondary quotation of the card).
Nemotron-Labs-Sigma 3B and 8B, plus distilled step variants
- Release: no announcement, no blog, no build.nvidia.com listing. The base Nemotron-Labs-Sigma-8B repo was published October 5, 2026 as a pretrained research model (per a card-reading analysis). The SFT 8B and the distilled variants followed in November 2026: the distill repo’s first commit is November 9, 2026 and its card dates the release November 12, 2026, the same day v2 of the paper landed.
- Paper: “Large Language Continuous Diffusion Models”, arXiv:2610.02665 (v2 November 12, 2026). Sigma is presented as the first family of large-scale continuous diffusion language models, at 3B and 8B.
- Architecture: Ministral-style decoder-only Transformer (34 layers, hidden 4,096, FFN 14,336, 32 query heads, 8 KV heads, SwiGLU, RoPE with YaRN) plus a 16-dimensional continuous token-embedding diffusion head; tokens are points in a 16-dim embedding space, noised with Gaussian noise and denoised in blocks. Warm-started from pretrained autoregressive weights rather than trained from scratch.
- Parameter counts: base 8B card name, 8.49B by safetensors census (8,489,947,136 BF16 plus about 2.1M F32). The distilled variants keep the auxiliary autoregressive readout head, so Nemotron-Labs-Sigma-distill-8B-nfe4 censuses at 9.03B. Each distill variant (distill-8B-nfe4 and siblings) is its own student checkpoint with the denoising step count baked in by training; the step count cannot be overridden at inference by design.
- Context: 4,096 tokens for prompt and output combined (both 8B variants, per card readings). The SFT repo lists 8,192 context in HF metadata; which figure governs which repo is a gap item.
- License: NVIDIA Nemotron Open Model License (card, per card-reading analysis). Offline use only; NVIDIA hosts no endpoint.
- Benchmarks: the base 8B card publishes NO benchmark table and points to the paper. The paper reports about 3.46x throughput over the model’s own autoregressive decoding on a single GB200 under SGLang at batch 1, about 8 tokens per forward on math and 4 on coding at negligible accuracy loss. The distill-8B-nfe4 card is the only Sigma card with a per-benchmark table: GSM8K, MATH (Minerva), HumanEval, HumanEval+, MBPP, MBPP+ at ten seeds, six-task mean 66.36. The same card flags its own figures as provisional (marked pending_ten_seed_six_task_validation; a single seed-42 run on three tasks), so treat them as preliminary by the card’s own instruction.
- Builder note: Sigma and Lightning share zero benchmarks on their cards. There is no published head-to-head; anyone claiming one is comparing across incompatible eval sets.
12. Public training material inventory
What a builder can actually download and study, by type.
Pretraining datasets
- Nemotron-CC (arXiv:2412.02595, December 2024): 6.3T tokens (4.4T deduplicated real plus 1.9T synthetic), with a 1.1T high-quality subset. Released under Common Crawl Terms of Use; the curation pipeline is Apache 2.0 NeMo Curator. Nemotron-CC-v2 and v2.1 extend it (9.1T tokens in the Super VL card’s corpus table).
- Nemotron-CC-Math-v1 (arXiv:2508.15096): 133B tokens at 3+ grade quality, 52B at 4+.
- Nemotron-CC-Code-v1: 427.9B tokens (Super VL card).
- Nemotron-Pretraining-Code v1 to v3: 1.7T tokens combined (Super VL card).
- Nemotron-Pretraining-Specialized v1.x and Nemotron-Pretraining-SFT-v1: 660B tokens combined (Super VL card).
- Nemotron-Pre-Training-Datasets collection (HF): the disclosed Nemotron 3 Nano corpus, 10,648,823,153,919 tokens total.
- Indic synthetic pretraining corpus: about 2.83T tokens across 22 languages, co-created with IIT Madras, released with Nemotron 3.5 Super VL.
- Granary-v2 (speech): 843,600 audio hours across 48 languages; public manifest about 463,062 hours. Released with Nemotron 3.5 Transcribe.
Post-training datasets
- HelpSteer2 (CC-BY-4.0): the preference dataset behind the 2024 RLHF wave.
- Llama-Nemotron-Post-Training-Dataset (HF): the full post-training data for the 2025 Llama-Nemotron reasoning family.
- Nemotron-Post-Training-Dataset-v1 and Nemotron-SFT-Data (18M+ samples per a secondary compilation of NVIDIA’s dataset listings).
- Nemotron-Super-Post-Training-Data: SFT datasets plus RL environment configurations for Nemotron 3 Super.
- Nemotron 3 Ultra Base releases: 10M SFT samples and 1M RL tasks under OpenMDW-1.1 (per a secondary summary of the Base card).
- AceReason-Math; OpenScienceReasoning-2; OpenCodeReasoning; Terminal-Corpus: the specialist lines’ datasets, all on HF.
- Nemotron-RL-Data: RL task data listed among NVIDIA’s HF dataset releases (secondary compilation).
Code and recipes
- NeMo, Megatron-LM, Megatron Bridge, NeMo RL, NeMo Gym, NeMo Curator, NeMo-Skills: open source frameworks covering pretraining, post-training, RL environments, data curation, and evaluation.
- NVIDIA-NeMo/Nemotron GitHub repo: hosts the Nemotron 3 family recipes per NVIDIA’s claims, and the Nemotron 3 Super tech report states its recipe is shared there.
- Nemotron-Labs-Diffusion: training recipe and code on GitHub, linked from the HF blog post.
- Explicit exception: Llama-Nemotron has open weights and open post-training data but NO published training recipes, per NVIDIA’s own Megatron Bridge documentation.
13. Size ladder (smallest to largest, by total parameters)
| Total | Model | Active | Type | Released |
|---|---|---|---|---|
| 1.5B | OpenReasoning-Nemotron-1.5B | 1.5B | Dense (Qwen2.5 derivative) | Jul 2025 |
| 1.5B | Nemotron-Research-Reasoning-Qwen-1.5B | 1.5B | Dense (Qwen derivative) | May 2025 |
| 3B | Nemotron-Labs-Diffusion-3B | 3B | Dense tri-mode diffusion (Ministral 3) | May 2026 |
| 3B | Nemotron-Labs-Sigma-3B | 3B | Dense continuous diffusion | Oct to Nov 2026 |
| 4B | Nemotron 3 Nano 4B | 4B | Dense hybrid Mamba | Mar 2026 |
| 4B | Nemotron-H-4B-Base-8K | 4B | Dense hybrid Mamba | 2025 |
| 4B | Llama-3.1-Minitron-4B | 4B | Dense (Llama derivative) | 2024 |
| 4B | Nemotron-Mini-4B / Mini-Hindi-4B | 4B | Dense (Nemotron-4 derivative) | 2024 |
| 4B | Llama-3.1-Nemotron-Nano-4B-v1.1 | 4B | Dense (Llama derivative) | May 2025 |
| 4.3B | Nemotron 3.5 Content Safety | 4.3B | Guard model | Mar 2026 |
| 7B | OpenMath-Nemotron-7B | 7B | Dense (Qwen derivative) | Apr 2025 |
| 7B | OpenCodeReasoning-Nemotron-7B | 7B | Dense (Qwen derivative) | Apr 2025 |
| 7B | AceReason-Nemotron-7B / 1.1-7B | 7B | Dense (Qwen derivative) | May to Jun 2025 |
| 7B | OpenReasoning-Nemotron-7B | 7B | Dense (Qwen derivative) | Jul 2025 |
| 8B | Nemotron-3 8B (2023) | 8B | Dense Transformer | Nov 2023 |
| 8B | Nemotron-H-8B-Base-8K | 8B | Dense hybrid Mamba | Mar 2025 |
| 8B | Llama-3.1-Nemotron-Nano-8B-v1 | 8B | Dense (Llama derivative) | Mar 2025 |
| 8B | Mistral-NeMo-Minitron-8B-Instruct | 8B | Dense (Mistral derivative) | Oct 2024 |
| 8B | Nemotron-Cascade-8B / 8B-Thinking | 8B | Dense (Qwen3 derivative) | Dec 2025 |
| 8B | Nemotron-Terminal-8B | 8B | Dense (Qwen3 derivative) | Feb 2026 |
| 8B | Nemotron-Labs-Diffusion-8B (+ 8B VLM) | 8B | Dense tri-mode diffusion | May 2026 |
| 8B | Nemotron-Labs-Sigma-8B (+ distill variants) | 8B (8.49B base, 9.03B distill census) | Dense continuous diffusion | Oct to Nov 2026 |
| 8B total / 4B eff. | Nemotron 3.5 Bolt E4B | 4B effective | Dense (Gemma 4 derivative) | Oct 2026 |
| 9B | Nemotron-Nano-9B-v2 | 9B | Hybrid Mamba (dense) | Aug 2025 |
| 12B | Nemotron-Nano-12B-v2-Base | 12B | Hybrid Mamba (dense) | Aug 2025 |
| 14B | AceReason-Nemotron-14B | 14B | Dense (Qwen derivative) | May 2025 |
| 14B | OpenMath / OpenCodeReasoning / OpenReasoning-14B | 14B | Dense (Qwen derivatives) | Apr to Jul 2025 |
| 14B | Nemotron-Cascade-14B-Thinking | 14B | Dense (Qwen3 derivative) | Dec 2025 |
| 14B | Nemotron-Terminal-14B | 14B | Dense (Qwen3 derivative) | Feb 2026 |
| 14B | Nemotron-Labs-Diffusion-14B | 14B | Dense tri-mode diffusion | May 2026 |
| 30B-A3B | Nemotron 3 Nano 30B-A3B (+ Omni) | 3.5B (Nano) | Hybrid Mamba MoE | Dec 2025 (Omni Apr 2026) |
| 30B-A3B | Nemotron-Cascade-2-30B-A3B | 3B | Hybrid Mamba MoE (post-train of Nano base) | Mar 2026 |
| 30B-A3B | Nemotron 3.5 Lightning 30B-A3B | about 3B | Hybrid Mamba MoE | Aug 2026 |
| 30B-A4B | Nemotron 3.5 Transcribe | 4.3B | Hybrid Mamba MoE speech | Oct 2026 |
| 32B | OpenMath / OpenCodeReasoning / OpenReasoning-32B | 32B | Dense (Qwen derivatives) | Apr to Jul 2025 |
| 32B | Nemotron-Terminal-32B | 32B | Dense (Qwen3 derivative) | Feb 2026 |
| 47B | Nemotron-H-47B-Base | 47B | Dense hybrid Mamba | 2025 |
| 49B | Llama-3.3-Nemotron-Super-49B-v1 / v1.5 | 49B | Dense (Llama derivative, NAS) | Mar / Jul 2025 |
| 51B | Llama-3.1-Nemotron-51B-Instruct | 51B | Dense (Llama derivative, NAS) | Sep 2024 |
| 56B | Nemotron-H-56B-Base | 56B | Dense hybrid Mamba | 2025 |
| 70B | Llama-3.1-Nemotron-70B-Instruct / Reward | 70B | Dense (Llama derivative) | Oct 2024 |
| 74B-A9B | Nemotron 3.5 Bolt 74B-A9B | 9B | Hybrid Mamba MoE (2-bit shipping artifact) | Oct 2026 |
| 120B-A12B | Nemotron 3 Super 120B-A12B | 12B | Hybrid Mamba LatentMoE, NVFP4 trained | Mar 2026 |
| 120B-A12B | Nemotron 3.5 Super VL 120B-A12B | 12B | Hybrid Mamba MoE + vision encoder | Oct 2026 |
| 253B | Llama-3.1-Nemotron-Ultra-253B-v1 | 253B | Dense (Llama derivative, NAS) | Apr 2025 |
| 340B | Nemotron-4-340B (Base / Instruct / Reward) | 340B | Dense Transformer | Jun 2024 |
| 550B-A55B | Nemotron 3 Ultra 550B-A55B (+ Base, GenRM) | 55B | Hybrid Mamba LatentMoE, NVFP4 trained | Jun 2026 |
| 550B-A55B | Nemotron-Labs-Teacher-Competition-Coding | 55B | Hybrid Mamba LatentMoE (Ultra teacher) | Aug 2026 |
Side lines not placed in the ladder: Nemotron 3.5 Embed 1B and 8B (secondary listing only); Nemotron-Parse-2.0 (size not verified); Nemotron 3.5 VoiceChat (size not verified).
14. Matched-size comparison candidates
For the masterclass comparison spine: same size band, different training and architecture choices, with enough public material to compare recipes and not just scores.
7B to 9B band (the densest band in the family)
- Nemotron-H-8B-Base-8K: NVIDIA’s own dense hybrid pretrain. The from-scratch data point.
- Nemotron-Nano-9B-v2: NVIDIA’s own hybrid, FP8-trained at 12B then Minitron-compressed to 9B. Compare against Nemotron-H-8B for what the extra pretraining tokens and the compression step each buy.
- Llama-3.1-Nemotron-Nano-8B-v1: no NVIDIA pretraining at all; Llama base plus SFT, RPO, RLVR. The clean post-training-only control.
- AceReason-Nemotron-7B and OpenReasoning-Nemotron-7B: same Qwen lineage, RL versus SFT-only. The pair isolates what GRPO adds over curated SFT at fixed size.
- Nemotron-Cascade-8B: Qwen3 base plus Cascade RL. Third post-training recipe at the same size.
- Nemotron-Terminal-8B: same band, narrow-domain SFT on synthetic task data; Terminal-Bench 2.0 gains show the dataset effect in isolation.
- Nemotron-Labs-Diffusion-8B and Nemotron-Labs-Sigma-8B: same Ministral-derived size class, two different diffusion objectives (masked tri-mode versus continuous). The architecture-axis comparison.
- Nemotron-3 8B (2023): historical control. Same nominal size, three years older, dense Transformer, 4K context.
30B-A3B band
- Nemotron 3 Nano 30B-A3B: the base recipe (25T tokens, SFT plus GRPO RL).
- Nemotron-Cascade-2-30B-A3B: the same base, post-trained with Cascade RL plus on-policy distillation. A controlled recipe comparison on identical weights lineage.
- Nemotron 3.5 Lightning 30B-A3B: the next iteration of the same slot, OpenMDW-1.1, tuned for harness throughput over peak scores.
- Nemotron 3 Nano Omni 30B-A3B: same skeleton, multimodal extension. Shows what the vision and audio towers cost and add.
- Nemotron 3.5 Transcribe 30B-A4B: the Lightning decoder reused as a speech backbone. Cross-modal reuse of one checkpoint family.
70B band
- Llama-3.1-Nemotron-70B-Instruct: RLHF (REINFORCE) on HelpSteer2, 2024 recipe.
- Llama-3.3-Nemotron-Super-49B-v1.5: nominally 49B, but it is the NAS-compressed descendant of the 70B class and is marketed against 70B models; include it here with the caveat stated.
- Nemotron 3.5 Bolt 74B-A9B: the 2026 answer at this size: MoE, 9B active, distilled and compressed from a 120B parent, shipped at 2 bits. The contrast with the dense 70B is the whole modern efficiency argument in one pair.
- Note: NVIDIA has no from-scratch dense 70B pretrain in the open roster. Every 70B-class entry is a Llama derivative or a compressed MoE. That absence is itself a finding.
100B+ band
- Nemotron-4-340B: dense, 9T tokens, 4K context, 2024. The old flagship recipe.
- Llama-3.1-Nemotron-Ultra-253B-v1: dense NAS derivative of a 405B, reasoning post-training, single-node serving target.
- Nemotron 3 Super 120B-A12B: hybrid LatentMoE, NVFP4 pretraining, 25T tokens, multi-environment RL. The recipe-modernization comparison against the 340B.
- Nemotron 3.5 Super VL 120B-A12B: Super plus vision, Ultra distillation, and Indic continued pretraining. Shows the 3.5 step as a delta on the Super recipe.
- Nemotron 3 Ultra 550B-A55B: the full stack: NVFP4, LatentMoE, MOPD from 10+ public teacher checkpoints. Compare its published data and teacher releases against every other lab’s flagship opacity.
15. Gaps and unverified items
Everything below was NOT confirmed at a primary source in this pass. Do not cite these as settled.
- Nemotron-H license conflict: HF metadata tags Nemotron-H-8B-Base-8K with the NVIDIA Internal Scientific Research and Development Model License, while redistributors cite the NVIDIA Open Model License. The card text itself was not read directly. This is the single most consequential open item in the roster, because it decides whether Nemotron-H is commercially usable.
- Nemotron 3 Super license: two secondary summaries of the card agree on NVIDIA Nemotron Open Model License, but the HF card was not read directly in this pass.
- Nemotron 3 Ultra pretraining tokens: the HF card says about 20T; a developer-blog-derived summary says a 10T base plus 212B new tokens. Card cited in the entry; conflict unresolved.
- Nemotron 3 Nano routing detail: the NGC card text states both “5 experts activated per token” and “top-6” in different lines, and one line says “2 shared experts” where the HF-side descriptions say 1 shared expert. The config file would settle it; it was not read.
- Nemotron 3 Nano active parameters: card headline 3.5B; secondary computations give 3.2B excluding embeddings. Both reported in the entry.
- Nemotron-4 15B: referenced as the compression base for the Mini-4B models, but an open weight release of the 15B itself was not confirmed.
- Nemotron-4-340B-Reward RewardBench figure: 92.0 and 94.1 both circulate in secondary sources. The tech report was not re-read for the exact table value.
- Nemotron-3 8B (2023) pretraining tokens: 3.5T per the NGC QA card, about 3.8T per a secondary compilation.
- Exact card license strings not pinned for: Nemotron-Nano-9B-v2 (reported NVIDIA Open Model License), Nemotron-Cascade-2-30B-A3B (reported NVIDIA Open Model License), Nemotron 3 Nano Omni (HF tag reported as nvidia-open-model-agreement), the Minitron derivatives, and the Qwen specialist lines where licenses are marked secondary in their entries.
- Nemotron-Terminal checkpoint licenses: not read. Qwen3 upstream is Apache 2.0, but the checkpoints’ own terms govern and were not verified.
- Nemotron-Labs-Sigma context: 4,096 tokens per card readings of the 8B repos; the SFT repo’s HF metadata lists 8,192. Which repos carry which limit was not resolved. Sigma 3B release dates and per-repo details rest on one card-reading analysis (OrcaRouter, November 17, 2026) plus the arXiv paper; NVIDIA published nothing.
- Jet-Nemotron: a small hybrid research line spotted in one secondary wiki snippet (August 2025). Not verified at any primary source; excluded from the roster body.
- Nemotron 3.5 side lines: Content Safety 4.3B, Embed 1B and 8B, VoiceChat, and Nemotron-Parse-2.0 rest on secondary catalogs and listings only. No cards were read.
- Nemotron 3 Super active parameters: 12B nominal in the tech report; 12.7B including embeddings per one analysis. Report figure cited.
- Llama-3.1-Nemotron-70B headline numbers (Arena Hard 85.0, MT-Bench 8.98) are quoted from the card via secondary sources; the card was not re-read in this pass.
- Nemotron-Cascade (first generation) license: not pinned in this pass; the research page confirms the data and models were released but the license name was not captured.