NVIDIA provides exemplary documentation for the Nemotron 3 Nano architecture. The model is explicitly described as a hybrid Mamba-2 and Transformer Mixture-of-Experts (MoE) model. Technical reports and white papers detail the specific layer composition (23 Mamba-2 layers, 23 MoE layers, and 6 GQA layers). The training methodology, including the use of the Warmup-Stable-Decay (WSD) learning rate schedule and the two-phase pre-training curriculum (diversity phase followed by high-quality phase), is thoroughly documented. The transition from previous generations (Nemotron-H and Nemotron 2 Nano) is also clearly explained.

The model was trained on a massive 25 trillion token corpus. NVIDIA discloses major data categories including web (Common Crawl), code (GitHub), math, and science. Specific datasets like Nemotron-CC-Code-v1 (427.92B tokens) and the InfiniByte cross-domain dataset are named. While exact percentage breakdowns for all 141 datasets are not provided in a single table, the technical report describes the curation process ('efficient data' paradigm) and the use of synthetic data generated via Lynx pipelines and LLM-based filtration (e.g., using Qwen3-30B-A3B for QA pair generation).

The tokenizer is publicly available via Hugging Face and the NeMo framework. It supports 20 languages and 43 programming languages, aligning with the model's claimed capabilities. Documentation specifies the use of special tokens for reasoning (<think> and </think> with IDs 12 and 13). While the exact vocabulary size is not prominently featured in marketing summaries, it is verifiable through the provided AutoTokenizer implementation and model configuration files.

NVIDIA is highly transparent regarding parameter density, explicitly distinguishing between total and active parameters. The model has 31.6B total parameters with ~3.2B active per forward pass (3.6B including embeddings). The MoE structure is detailed as having 128 routed experts plus 1 shared expert per layer, with a routing mechanism that activates 6 experts per token. This level of detail prevents the common 'parameter inflation' seen in other MoE disclosures.

While NVIDIA specifies the hardware used for inference (H100, A100, H200) and the software framework (Megatron-LM, NeMo), it provides limited information on the total training compute budget in terms of GPU-hours. The training duration (September to December 2025) is known, and the batch size (3072) is disclosed, but a formal carbon footprint calculation or total cost estimate is conspicuously absent from the public technical reports.

Reproducibility is a core focus of the Nemotron 3 release. NVIDIA provides the NeMo Evaluator SDK, which includes the exact YAML configurations, prompt templates, and sampling parameters used for the model card benchmarks. Evaluation results are compared against peers (Qwen3, GPT-OSS) with specified versions. The disclosure of 'Reasoning ON/OFF' modes and their impact on benchmark scores (e.g., AIME 2025 with and without tools) demonstrates a high commitment to verifiable performance claims.

The model exhibits strong identity consistency, correctly identifying its version and capabilities (such as its 1M token context window and reasoning traces). It does not attempt to mimic competitor identities and is transparent about its 'Thinking' budget and configurable reasoning depth. The distinction between the 'Base' and 'Instruct' (post-trained) versions is clearly maintained across all documentation.

The model is released under the 'NVIDIA Open Model License'. This is a custom permissive license that explicitly allows commercial use and the creation of derivative works. However, it is not a standard OSI-approved license like Apache 2.0 or MIT, and it contains specific clauses regarding NVIDIA's 'Trustworthy AI' terms. While clear, the use of a proprietary license instead of a standard open-source one introduces some legal complexity for users.

Hardware requirements are well-documented for various configurations. NVIDIA provides VRAM estimates for FP16, INT8, and INT4 (e.g., ~62GB for FP16 at 1k context). It explicitly warns about the memory scaling of the 1M token context window, noting that 24GB consumer cards (like the RTX 4090) can run the model but may crash if the context is set to the full 1M limit without significant quantization. Support for FP8 and its accuracy trade-offs (99% recovery) is also documented.

NVIDIA uses a clear naming convention (Nemotron 3 Nano 30B-A3B) and provides data cutoff dates (June 2025 for pre-training, November 2025 for post-training). While a formal semantic versioning changelog for weight updates is not as robust as software versioning, the release of specific checkpoints (BF16, FP8) and the integration with the NeMo framework versioning (e.g., NeMo 25.11.01) provides a reasonable track for developers.