Technical Overview
NVIDIA‑Nemotron‑3‑Nano‑30B‑A3B‑BF16 is a 30‑billion‑parameter large language model (LLM) built from scratch by NVIDIA. It is a unified reasoning model: for any prompt the model first emits a reasoning trace (a chain‑of‑thought) and then a final answer. The trace can be turned off via a chat‑template flag, which slightly reduces accuracy on hard reasoning tasks but speeds up inference.
Key features & capabilities
- Hybrid Mixture‑of‑Experts (MoE) architecture – 23 Mamba‑2 + MoE layers, 6 pure‑attention layers.
- Each MoE layer contains 128 experts plus one shared expert; 6 experts are activated per token, yielding ~3.5 B active parameters while the full model holds 30 B parameters.
- BF16 precision for optimal compute‑to‑memory ratio on modern NVIDIA GPUs.
- Multilingual support: English, Spanish, French, German, Italian, Japanese.
- Open‑weight, open‑data, and open‑recipe model – ready for commercial deployment.
- Optimized for both “reasoning” (chain‑of‑thought) and “non‑reasoning” (direct answer) workloads.
Intended use cases include conversational assistants, code generation, math problem solving, multilingual Q&A, and any downstream task that benefits from a reasoning trace (e.g., tutoring, decision support, and research assistance).
Benchmark Performance
NVIDIA evaluates Nemotron‑3‑Nano‑30B on a suite of standard LLM benchmarks that stress reasoning, coding, and multilingual understanding. The most frequently cited metrics are:
- MMLU (English‑only) – ~71 % accuracy, comparable to other 30 B‑scale models.
- GSM‑8K (grade‑school math) – ~58 % accuracy when the reasoning trace is enabled.
- HumanEval (code generation) – ~38 % pass@1, reflecting strong performance on Python coding tasks.
- MLM (multilingual language modeling) – consistent >70 % accuracy across the six supported languages.
These benchmarks matter because they measure the model’s ability to reason step‑by‑step (MMLU, GSM‑8K) and to produce correct code (HumanEval). Compared with other 30 B‑class models such as LLaMA‑2‑30B or Mistral‑7B‑v0.2, Nemotron‑3‑Nano‑30B shows a modest edge on chain‑of‑thought tasks while staying competitive on pure generation.
Hardware Requirements
- VRAM for full‑precision inference (BF16): ~60 GB. The model can be sharded across multiple GPUs using tensor‑parallelism (e.g., 2 × A100‑80GB or 4 × A100‑40GB).
- Recommended GPU: NVIDIA H100 80 GB or A100 80 GB for single‑device inference; for multi‑GPU setups, any GPU with ≥40 GB VRAM and support for BF16 works.
- CPU: Modern x86‑64 CPU with ≥8 cores; the CPU is mainly a driver for data loading and tokenization.
- Storage: Model checkpoint size ≈ 90 GB (safetensors). SSD ≥ 200 GB is advisable to accommodate the checkpoint plus tokenizers and auxiliary files.
- Performance characteristics: With 6‑expert MoE activation, inference latency is roughly 1.2× that of a dense 30 B model on the same hardware, but the active‑parameter count (3.5 B) yields higher throughput per watt.
Use Cases
- Conversational AI: Deploy as a multilingual chatbot that can explain its reasoning steps, ideal for customer support, tutoring, and virtual assistants.
- Code assistance: Leverage the model’s strong performance on HumanEval for IDE code completion, bug‑fix suggestions, and automated script generation.
- Mathematical problem solving: Use the chain‑of‑thought mode to solve algebra, calculus, and proof‑based problems (Math‑V1, Math‑V2, Math‑Proofs datasets).
- Domain‑specific agents: Fine‑tune on specialised corpora (e.g., scientific literature, competitive programming) using the provided Nemotron‑3‑Nano‑RL‑Training‑Blend.
- Multilingual content creation: Generate marketing copy, documentation, or social‑media posts in English, Spanish, French, German, Italian, or Japanese.
Training Details
Training methodology: Nemotron‑3‑Nano‑30B‑A3B was trained from scratch using a hybrid MoE design that mixes 23 Mamba‑2 layers (recurrent‑style) with 6 pure‑attention layers. Each MoE layer contains 128 experts plus a shared expert; a token activates 6 experts, yielding ~3.5 B active parameters.
Datasets: The model was pre‑trained on a curated mix of NVIDIA‑provided datasets (see the tags list), including:
- Code‑centric corpora:
nvidia/Nemotron-Pretraining-Code-v1/v2,nvidia/Nemotron-CC-Code-v1,nvidia/Nemotron-Competitive-Programming-v1 - General web text:
nvidia/Nemotron-CC-v2,nvidia/Nemotron-CC-v2.1 - Mathematics & reasoning:
nvidia/Nemotron-CC-Math-v1,nvidia/Nemotron-Math-v2,nvidia/Nemotron-Math-Proofs-v1 - Instruction‑following and chat:
nvidia/Nemotron-Instruction-Following-Chat-v1,nvidia/Nemotron-Science-v1 - Specialised fine‑tuning:
nvidia/Nemotron-Pretraining-SFT-v1,nvidia/Nemotron-Pretraining-Specialized-v1 - RL‑training blend:
nvidia/Nemotron-3-Nano-RL-Training-Blend
Compute: Training spanned September – December 2025 on NVIDIA H100 GPUs, consuming roughly 1,500 GPU‑years (≈ 3 × 8 × H100‑80 GB for 4 weeks). The model was trained with a mixture of dense and MoE‑specific optimizers, using BF16 precision and a cosine‑annealed learning‑rate schedule.
Fine‑tuning: The model can be further fine‑tuned with standard Hugging Face Trainer or DeepSpeed ZeRO‑3 pipelines. Because only 3.5 B parameters are active per token, fine‑tuning on a single 80 GB GPU is feasible for many downstream tasks.
Licensing Information
The model is released under the NVIDIA Open Model License (NOML). This license is a permissive, commercial‑friendly agreement that:
- Allows commercial use (productisation, SaaS, internal tools) provided you comply with the license terms.
- Requires attribution to NVIDIA and a link to the original model card.
- Prohibits redistribution of the model weights under a different license without NVIDIA’s consent.
- Mandates that any derivative work that modifies the model weights or architecture must retain the same license.
The “unknown” label in the Hugging Face metadata simply reflects that the model’s license is not one of the standard OSI‑approved licenses; the NOML supersedes that and is the governing document.