31.58B total; approximately 3.6B active including embeddings (3.2B under NVIDIA’s embedding-excluded convention); Base and post-trained checkpoints, with BF16 and post-trained FP8 packaging
Post-trained: up to 1M tokens in report/current support, 262,144 configured default, but pinned launch BF16 card declares 128K input; Base: 128K declared input/output versus 262,144 configuration and launch retrieval tests through 512K; family-level 1M scope unresolved for Base
Custom NVIDIA Open Model License Agreement at launch (October 24, 2025 version, including Trustworthy AI and guardrail conditions); subsequently NVIDIA Nemotron Open Model License (December 15, 2025 version); commercial use permitted, not Apache-2.0
Text-only, from-scratch hybrid Mamba-2–Transformer MoE model for reasoning, coding, tool use and agentic workflows. Its 52 sequential residual blocks comprise 23 Mamba-2, 23 MoE and 6 grouped-query attention blocks (32 query heads, 2 KV heads), not parallel Mamba/attention branches. Each MoE block routes to 6 of 128 experts and adds one shared MLP of twice the routed-expert width; no exported multi-token-prediction module. Base pretraining uses 25T processed tokens in a two-phase quality curriculum, followed by 121B tokens of long-context continued pretraining mixing 512K and 4K sequences; the disclosed corpus contains about 13.34T tokens, with more than 3T new unique tokens over Nemotron 2. The post-trained checkpoint adds supervised fine-tuning over 18M samples with 256K packing, synchronous multi-environment GRPO for verifiable math, code, science, instruction following and tool-use rewards, then GenRM-guided RLHF with group-relative length control. Synthetic teacher data includes Qwen, GPT-OSS and DeepSeek-generated traces; the weights are not a compressed Qwen checkpoint. DPO tool-hallucination experiments in the report were not used in the released model. Reasoning on/off is controlled through chat-template prefixes; finite reasoning budgets require generation-side stopping/continuation logic rather than a hard checkpoint switch. Supports English, German, Spanish, French, Italian and Japanese in post-training. Reported BF16 results include AIME25 89.1 without tools versus 99.2 with tools, GPQA 73.0 without tools, LiveCodeBench v6 68.3, SWE-Bench 38.8 with OpenHands, BFCL v4 53.8 and RULER-100 86.3 at 1M with reasoning off; long-document AA-LCR 35.9 and multilingual MMLU-ProX 59.5 trail Qwen3-30B-A3B-Thinking-2507. Evaluation uses benchmark-specific scaffolds and repeated sampling, not universally single-attempt standalone answers. NVIDIA reports 3.3× throughput versus Qwen3-30B-A3B-Thinking-2507 and 2.2× versus GPT-OSS-20B on one H200 with 8K input/16K output, taking the better vLLM/TRT-LLM configuration; Nemotron and Qwen use FP8, GPT-OSS MXFP4. Selective post-training FP8 quantization retains sensitive attention and preceding Mamba blocks in BF16 and approximately 99% median benchmark accuracy recovery. Original Base-BF16, post-trained BF16 and post-trained FP8 are grouped; later NVFP4, Nano 4B, Omni, Super and Ultra releases are excluded.