NVIDIA Debuts Nemotron 3 Family of Open Models

TL;DR

NVIDIA’s December-2025 Nemotron 3 announcement positions the family as the umbrella open release for the post-CES 2026 cycle, formalizing the Nano / Super / Ultra size tier, a “hybrid latent MoE” architecture, and a coordinated drop of 3T pretraining + post-training tokens, the NeMo Gym / NeMo RL training-environment libraries, and a Nemotron Agentic Safety Dataset. Only Nemotron 3 Nano (30B-total / 3B-active, 1M context) is available at announcement; Super (~100B / ~10B-active) and Ultra (~500B / ~50B-active) are slated for the first half of 2026, both trained in NVFP4 on Blackwell. The headline efficiency claim is 4× higher token throughput than Nemotron 2 Nano and up to 60% reduction in reasoning-token generation, and the model card claims #1 ranking on Artificial Analysis among same-size models. Significant as the parent-level release that supersedes the per-product Nemotron 3 Super / Nano-Omni / Cascade / Labs-Diffusion / Jet-Nemotron papers already filed in the wiki.

Key claims

• Nemotron 3 is announced as a three-tier family — Nemotron 3 Nano (30B total / 3B active, 1M-token context), Super (~100B / ~10B active), and Ultra (~500B / ~50B active) — using a “hybrid latent mixture-of-experts” architecture for multi-agent workloads [News Summary, §Nemotron 3 Reinvents Multi-Agent AI].
• Nemotron 3 Nano claims 4× higher token throughput than Nemotron 2 Nano and up to 60% reduction in reasoning-token generation [§Nemotron 3 Reinvents Multi-Agent AI].
• Artificial Analysis ranks Nemotron 3 Nano as the most open and efficient model at its size, with leading accuracy among the comparison set [§Nemotron 3 Reinvents Multi-Agent AI, Artificial Analysis link].
• Super and Ultra are trained in 4-bit NVFP4 on Blackwell, claimed to reduce memory and speed up training without accuracy loss relative to higher-precision formats [§Nemotron 3 Reinvents Multi-Agent AI].
• The release bundles 3T tokens of new pretraining + post-training datasets, the NeMo Gym training-environments library, NeMo RL, and the Nemotron Agentic Safety Dataset, all on Hugging Face and GitHub [§New Open Tools and Data for AI Agent Customization].
• Day-0 serving is supported across LM Studio, llama.cpp, SGLang, and vLLM, with hosted inference on Baseten, DeepInfra, Fireworks, FriendliAI, OpenRouter, and Together AI; cloud availability spans Amazon Bedrock plus “soon” on Google Cloud, CoreWeave, Crusoe, Microsoft Foundry, Nebius, Nscale, and Yotta [§Get Started With NVIDIA Open Models].
• Prime Intellect and Unsloth are integrating the NeMo Gym training environments directly into their workflows [§New Open Tools and Data for AI Agent Customization].
• Super and Ultra are not released at announcement — both are expected in H1 2026 [§Get Started With NVIDIA Open Models].

Method

This is an announcement post, not a technical report — the architecture is named “hybrid latent mixture-of-experts” but the actual design is linked out to the developer-blog tear-down at developer.nvidia.com/blog/inside-nvidia-nemotron-3-techniques-tools-and-data-that-make-it-efficient-and-accurate, which the announcement does not summarize. The three models follow a clean ~10× active-parameter scaling (3B → 10B → 50B) at ~5× total- parameter scaling (30B → 100B → 500B), implying a fixed-ratio sparse MoE within the family rather than per-tier topology changes. Super and Ultra are trained in NVFP4 on Blackwell; Nano’s precision is not stated in the post.

The release ships with three data/tooling primitives a downstream team would actually need:

• Pretraining data: 3T tokens across the nemotron-pre-training-datasets HF collection (huggingface.co/collections/nvidia/nemotron-pre-training-datasets).
• Post-training data: the nemotron-post-training-v3 HF collection (huggingface.co/collections/nvidia/nemotron-post-training-v3) covering reasoning, coding, multistep workflow examples.
• RL environments + framework: the NeMo Gym and NeMo RL open-source libraries, plus a safety dataset (Nemotron-AIQ-Agentic-Safety-Dataset-1.0).

Results

The headline numbers are all about Nano (Super and Ultra are unreleased):

• Throughput: 4× higher token throughput than Nemotron 2 Nano.
• Reasoning efficiency: up to 60% reduction in reasoning-token generation.
• Context: 1M-token context window.
• Independent ranking: Artificial Analysis ranks Nano #1 among same-size open models on the announcement-day intelligence-vs-output-speed view, against gpt-oss-20B, Llama-4 Maverick / Scout, Magistral Medium 2509, Olmo-3-32B Think, Apriel-V1.6-15B Thinker, Qwen3-30B-A3B 2507 Reasoning, and earlier Nemotron Nano variants.

No standalone benchmark table is reproduced in the post; comparison numbers against frontier closed models are absent.

Why it’s interesting

This is the umbrella announcement that ties together the family of Nemotron 3 papers and pre-releases already filed in the wiki — Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning for the Super tier, Nemotron 3 Nano Omni: Efficient and Open Multimodal Intelligence for the multimodal Nano variant, Nemotron-Cascade 2: Post-Training LLMs with Cascade RL and Multi-Domain On-Policy Distillation for the post-training recipe, Nemotron-Labs-Diffusion: A Tri-Mode Language Model Unifying Autoregressive, Diffusion, and Self-Speculation Decoding for the diffusion-decoding sibling, and Jet-Nemotron: Efficient Language Model with Post Neural Architecture Search for the architecture-search predecessor — confirming that all of these land under a single product brand with a coordinated data/tooling drop. It also extends the NVIDIA Unveils New Open Models, Data and Tools to Advance AI Across Every Industry CES bundle pattern from a multi-domain spread (Nemotron + Cosmos + Alpamayo + GR00T + Clara) to a deeper vertical drop — the new NeMo Gym + NeMo RL bring NVIDIA into the same managed-RL-environment territory as OpenReward — 330+ RL Environments through one API (General Reasoning announcement) (OpenReward) and SETA: Scaling Environments for Terminal Agents (env repo) (SETA), making this the first major-vendor entry on RL Environment Platforms. The 60%-reasoning-token-reduction claim parallels efficient-reasoning bets like JustRL: Scaling a 1.5B LLM with a Simple RL Recipe from a very different angle.

See also

• Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning — Super-tier technical report (hybrid Mamba-Transformer MoE for agentic reasoning).
• Nemotron 3 Nano Omni: Efficient and Open Multimodal Intelligence — multimodal Nano variant within the same family.
• Nemotron-Cascade 2: Post-Training LLMs with Cascade RL and Multi-Domain On-Policy Distillation — post-training recipe used across Nemotron tiers.
• Nemotron-Labs-Diffusion: A Tri-Mode Language Model Unifying Autoregressive, Diffusion, and Self-Speculation Decoding — diffusion-decoding Nemotron-Labs sibling.
• Jet-Nemotron: Efficient Language Model with Post Neural Architecture Search — predecessor architecture-search work on the efficient-Nemotron lineage.
• NVIDIA Unveils New Open Models, Data and Tools to Advance AI Across Every Industry — the broader CES 2026 multi-domain NVIDIA open release.
• Open foundation-model releases — concept page tracking coordinated open-weights drops.
• RL Environment Platforms — concept page tracking managed RL-environment libraries (NeMo Gym joins this set).
• MoE Routing Design — concept page where the “hybrid latent MoE” architecture will be a relevant datapoint once the developer-blog tear-down is filed.
• Synthetic Training Data — concept page; the 3T-token post-training dataset belongs in this lineage of large-scale synthesis/curation releases.