NVIDIA Nemotron v3

NVIDIA Nemotron 3 Ultra 550B A55B BF16 is a frontier-scale Mixture-of-Experts reasoning model designed for advanced agentic workflows, long-context analysis, and high-accuracy reasoning across code, mathematics, science, and multilingual tasks. It features 550B total parameters with 55B active, utilizing a hybrid LatentMoE architecture that combines Mamba-2, MoE, and attention layers with multi-token prediction for efficient inference. The model uses an 8,192 hidden size and 64 attention heads, activating 22 experts per token across 512 routed experts and a shared expert. Supporting up to a 1M token context window, it is optimized for complex reasoning, tool use, large-scale document analysis, and high-stakes retrieval-augmented generation workloads.

Qwen3 Nemotron 235B A22B GenRM is a large-scale Generative Reward Model (GenRM) built on the Qwen3-235B-A22B foundation and fine-tuned to evaluate assistant responses for helpfulness and quality. Based on a 235B-parameter Mixture-of-Experts (MoE) transformer, it features 94 layers, 64 attention heads, 128 experts (8 experts per token), and a 4,096 hidden size. The model processes up to 131K tokens and outputs structured helpfulness and ranking scores for candidate responses. It is designed for reinforcement learning from human feedback (RLHF), large-scale preference modeling, and advanced AI alignment workflows.

Nemotron 3.5 Super VL 120B A12B is a multimodal hybrid Mixture-of-Experts model built for reasoning, coding, tool use, and document and video understanding. The model has 120B total parameters and 12B active, spread across 88 layers with 4,096 hidden size, 32 attention heads, and 2 KV heads. Its stack interleaves Mamba-2, LatentMoE, and attention layers, with each MoE layer activating 22 of 512 experts plus a shared expert. A C-RADIO vision encoder processes images and video within a 1M-token context, while Multi-Token Prediction enables native speculative decoding.

Nemotron 3 Super 120B A12B BF16 is a hybrid Mixture-of-Experts reasoning model developed by NVIDIA for large-scale agentic workflows and long-context reasoning. It features a 120B parameter architecture with ~12B active parameters, built on an 88-layer hybrid backbone combining Mamba-2 sequence layers, Transformer attention, and Latent MoE routing with 512 experts. The model incorporates Multi-Token Prediction (MTP) for faster generation and supports up to a 1M token context window, enabling efficient long-horizon reasoning, multi-agent coordination, and large-scale retrieval tasks.

Nemotron 3.5 Bolt 74B A9B is a hybrid vision-language Mixture-of-Experts model designed for local agentic tasks, coding, and visual reasoning. It features 74B total parameters with 9B active, using an 88-layer architecture with 4,096 hidden size, 32 attention heads, and 2 KV heads. The backbone interleaves Mamba-2, LatentMoE, and attention layers, with each MoE layer activating 9 routed experts and a shared expert per token. Compressed from Nemotron 3.5 Super VL, it adds a C-RADIO v4-H vision encoder and 256K-token context, with the BF16 checkpoint aimed at post-training.

NVIDIA Nemotron 3.5 Lightning 30B A3B is a 30B-parameter hybrid Mixture-of-Experts model with 3B active parameters, combining Mamba-2, MoE, and Transformer attention layers. It features 52 layers, 128 routed experts with 6 activated per token, 32 attention heads, and a 2-head KV configuration, supporting up to 1M-token context, with multi-token prediction and reinforcement learning for reasoning, coding, tool use, and agentic workflows.

Nvidia Nemotron 3 Nano Omni 30B A3B Reasoning BF16 is a multimodal Mixture-of-Experts reasoning model designed for unified audio, vision, and language understanding, enabling AI agents to efficiently derive insights from videos and documents with high accuracy. It features a 30B parameter architecture with 3B activated, using 52 layers, 2,688 hidden size, and 32 attention heads, activating 6 experts per token across 128 routed experts with a shared expert. The model integrates C-RADIOv4-H based vision and Parakeet audio encoders, and combines hybrid Mamba-attention layers for efficient long-context processing and precise reasoning, supporting up to 256K tokens for image, video, document, and speech reasoning tasks.

Nemotron 3 Nano 30B A3B BF16 is a hybrid Mixture-of-Experts (MoE) large language model built for both reasoning-intensive and general-purpose AI workloads. The model contains 30B total parameters with 3.5B active parameters, activating 6 experts per token across 128 routed experts and 1 shared expert. Architecturally, it features 52 layers combining MoE, Mamba-2 state space layers, and grouped query attention, with 32 attention heads and a 2,688 hidden size. Supporting up to a 1M token context window, it is optimized for configurable reasoning, agentic workflows, and commercial AI deployments.

Nemotron 3 Nano 30B A3B Base BF16 is a hybrid Mixture-of-Experts (MoE) large language model trained from scratch for adaptable large-scale language modeling. The model contains 30B total parameters with 3.5B active parameters, activating 6 experts per token across 128 routed experts and 1 shared expert. It features a 52-layer architecture combining MoE and Mamba-2 state space layers with grouped query attention, 32 attention heads, and a 2,688 hidden size. Supporting up to a 1M token context window, this base checkpoint is designed for fine-tuning, domain adaptation, and building specialized AI systems.