Nemotron is NVIDIA's family of open models, released with open weights, training recipes, and those training datasets NVIDIA holds redistribution rights to, and aimed at agentic applications. NVIDIA describes the family as "highly efficient, multimodal, open AI models built for long-running, self-evolving agents" and states that Nemotron models are not merely open-weight but open source, with training data as well as weights published on Hugging Face for evaluation before deployment (Source: nvidia.com; developer.nvidia.com). The current generation is Nemotron 3, announced in December 2025 in Nano, Super and Ultra sizes, extended in August 2026 by the Nemotron 3.5 Lightning point release.
Family and release history
NVIDIA announced the Nemotron 3 family on 15 December 2025, describing three models intended for different points on the cost-accuracy curve: Nano, a 30-billion-parameter model activating up to 3 billion parameters at a time for targeted, efficient tasks; Super, a reasoning model of approximately 100 billion parameters with up to 10 billion active per token, for multi-agent applications; and Ultra, given at announcement as approximately 500 billion parameters with up to 50 billion active per token, for complex applications. The models as released came in slightly above those announcement figures — Nano at 31.6 billion total and 3.2 billion active, Ultra at 550 billion total and 55 billion active. Nano was released first, with Super and Ultra expected in the first half of 2026 (Source: nvidianews.nvidia.com; research.nvidia.com). NVIDIA characterises Super as optimised for collaborative agents and high-volume workloads such as IT ticket automation, and Ultra as the accuracy and reasoning ceiling of the family.
Nemotron 3 Ultra was announced at GTC San Jose in 2026 and published on NVIDIA's research site on 4 June 2026, with the model weights and the training recipe released alongside (Source: research.nvidia.com; github.com). Reporting circulated at the time placed the Ultra release "around June 2, 2026"; NVIDIA's own publication date is 4 June 2026 (Source: nlp.elvissaravia.com).
Nemotron 3.5 Lightning followed on 11 August 2026, released as NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 in a 1.0-preview version on Hugging Face and build.nvidia.com (Source: build.nvidia.com).
NVIDIA's investment in a further in-house open-source model family it calls Nemotron 4 became public on 11 August 2026 through an item in The Information; no article body was retrievable, and the model's scale, timing and licensing are unverified (Source: New Developments Log/2026-08-11-0812-ai-developments.md).
Architecture and training
The Nemotron 3 models share a hybrid Mamba-Transformer mixture-of-experts architecture, which NVIDIA presents as delivering higher throughput than a standard Transformer at better or comparable accuracy. Super and Ultra add three further techniques: LatentMoE, a hardware-aware expert design; Multi-Token Prediction layers, for long-form generation efficiency and native speculative decoding; and pre-training in NVFP4, NVIDIA's four-bit floating-point format. All Nemotron 3 models support context lengths up to one million tokens, are post-trained across a diverse set of reinforcement-learning environments, and expose granular reasoning-budget control at inference time (Source: research.nvidia.com).
Nemotron 3 Ultra has 550 billion total and 55 billion active parameters. Its post-training pipeline combines supervised fine-tuning, reinforcement learning, and multi-teacher on-policy distillation (Source: research.nvidia.com). Nemotron 3 Nano is 31.6 billion total parameters with 3.2 billion active (3.6 billion including embeddings), and achieves higher accuracy than the previous-generation Nemotron 2 Nano while activating less than half the parameters per forward pass (Source: research.nvidia.com).
For Nemotron 3.5 Lightning, NVIDIA representatives said the company used distillation to give the model capabilities similar to the larger Nemotron models, and described it as lightweight enough to run on a single graphics processing unit in a laptop or desktop (Source: cnbc.com). See Distillation.
Published training data
NVIDIA publishes the pre-training and post-training datasets it holds redistribution rights to. Those released with Nemotron 3 Nano include Nemotron-CC-v2.1, 2.5 trillion new English tokens from Common Crawl with synthetic rephrasing and translation; Nemotron-CC-Code-v1, 428 billion code tokens processed from Common Crawl code pages; Nemotron-Pretraining-Code-v2, a refreshed and deduplicated set of curated GitHub code references with large-scale synthetic code data; and Nemotron-Pretraining-Specialized-v1, synthetic datasets for areas such as STEM reasoning and scientific coding. Datasets released with Ultra include Nemotron-Pretraining-Code-v3, 173 billion tokens of GitHub code through 30 September 2025; Nemotron-Pretraining-Legal-v1, synthetic data intended to improve legal capabilities; Nemotron-Pretraining-Specialized-v1.2, covering factual recall, moral scenarios and generative and multiple-choice questions; and Nemotron-Posttraining-v3 for supervised fine-tuning and reinforcement learning (Source: research.nvidia.com; research.nvidia.com; github.com). The earlier Nemotron-CC-v2 (6.6 trillion quality-filtered Common Crawl tokens with multilingual question-answer data) and the Llama-Nemotron post-training dataset (2.2 million math and 500,000 code items) remain published for prior generations (Source: github.com).
Capabilities and benchmarks
NVIDIA's published comparisons emphasise inference throughput: the claims are that Nemotron matches open-weight peers on accuracy while running substantially faster. The figures below are as NVIDIA states them, including its own spelling of peer model names.
| Model | NVIDIA's reported comparison | Setting |
|---|---|---|
| Nemotron 3 Nano | 3.3× the inference throughput of Qwen3-30B-A3B; 2.2× that of GPT-OSS-20B | 8K input / 16K output, single H200 |
| Nemotron 3 Nano | Higher accuracy than GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507 | "popular benchmarks spanning different categories" |
| Nemotron 3 Nano | Outperforms GPT-OSS-20B and Qwen3-30B-A3B-Instruct-2507 on RULER across context lengths | up to 1M tokens |
| Nemotron 3 Ultra | 5.9× the inference throughput of GLM-5.1-754B-A40B; 4.8× that of Kimi-K2.6-1T-A32B; 1.6× that of Qwen-3.5-397B-17B | 8K input / 64K output |
| Nemotron 3 Ultra | "On-par accuracies compared to other state-of-the-art open LLMs" | diverse benchmark set |
| Nemotron 3 Ultra | Outperforms state-of-the-art open LLMs on RULER at 1M context | 1M tokens |
| Nemotron 3.5 Lightning | 4× higher throughput and 30% faster task completion than other leading open models of similar size | not stated |
(Source: research.nvidia.com; research.nvidia.com; ollama.com)
All of these are NVIDIA's own figures, published without independent replication. Third-party results reported elsewhere are narrower: Thinking Machines states that Inkling matches Nemotron 3 Ultra on Terminal Bench 2.1 at roughly one third of the tokens (Source: thinkingmachines.ai).
Availability
Nemotron model weights, including base, post-trained, and quantised checkpoints, are published on Hugging Face. Nemotron 3 Ultra is distributed in NVFP4, BF16 and base BF16 forms alongside the generative reward model used for reinforcement learning from human feedback; Nemotron 3 Nano in FP8, BF16 and base BF16 forms. Training recipes are published in the NVIDIA Nemotron developer repository on GitHub, and models are hosted for trial on build.nvidia.com (Source: research.nvidia.com; research.nvidia.com; github.com).
Nemotron 3.5 Lightning became available on the day of release through Ollama and Oracle Cloud Infrastructure Generative AI, and was launched alongside NeMo Switchyard, an NVIDIA model-routing component (Source: ollama.com; docs.oracle.com; blogs.nvidia.com).
Documentation and safety
NVIDIA publishes a family white paper, NVIDIA Nemotron 3: Efficient and Open Intelligence, and per-model technical reports for Nano, Super and Ultra (Source: research.nvidia.com). No system card, responsible-scaling commitment, or published safety case accompanies the family, and no third-party safety evaluation of a Nemotron model has been identified — a documentation posture that differs from the frontier-lab practice described at Frontier AI Governance.
Reception
Within the open-weight field, Nemotron appears most often as a comparison baseline rather than as a subject in its own right. Inkling Small's published comparison set places the Nemotron 3 family alongside a predominantly Chinese group of open-weight models — Qwen3.5-397B-A17B, MiMo V2.5, Minimax M2.7, DeepSeek V4 Flash, Kimi K3 and GLM-5.2 (Source: thinkingmachines.ai). A June 2026 survey of twelve open-weight models listed Nemotron 3 Super among the field alongside GLM-5.1, DeepSeek V4, Qwen3 and Gemma 4 (Source: blog.bytebytego.com). Nathan Lambert traces the Gated DeltaNet lineage behind Kimi K3's Kimi Delta Attention through architectures adopted by Qwen and Nemotron (Kimi K3: The open-weights escalation (Nathan Lambert, July 2026)). See Open-Weight Frontier Models and Open-Source AI / Open-Weight Models.
Relationships
- instance-of: Open-Source AI / Open-Weight Models, Open-Weight Frontier Models — NVIDIA publishes training data and recipes alongside weights, which it describes as open source rather than merely open-weight.
- depends-on: Distillation — multi-teacher on-policy distillation in Ultra's post-training; distillation from larger Nemotron models for 3.5 Lightning.
- related: Nvidia & TSMC — AI Compute Infrastructure — developer.
- related: Gemma (Google open-weight models), Inkling, Inkling-Small, Kimi K3, GLM-5.2, DeepSeek V4 Pro / V4 Flash — peer open-weight models and named comparison baselines.
Open questions
- NVIDIA's accuracy and throughput claims are self-published; no independent benchmark replication of Nemotron 3 Ultra or 3.5 Lightning has been identified.
- The scale, timing and licensing of the reported Nemotron 4 family are unverified.
- The date of Nemotron 3 Super's release is not established by the sources consulted. NVIDIA publishes a Super technical report and a June 2026 third-party survey lists Super among available open-weight models, but no dated release announcement was located.
Provenance note: No Nemotron technical report or white paper has been ingested; citations rest on NVIDIA's own pages and are marked
(Source: URL). Upgrade them to asources/wikilink if the Nemotron 3 white paper or a technical report is ingested.