nvidia/NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-BF16
NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-BF16 is a 30 billion parameter 3-in-1 elastic large language model developed by NVIDIA, featuring nested 30B, 23B, and 12B variants within a single BF16 checkpoint. Derived from Nemotron-3-Nano-30B, it utilizes a hybrid Mamba-Transformer-MoE architecture and the Elastic post-training framework to deliver efficient budget control and strong reasoning performance. This model is optimized for AI Agent systems, chatbots, and RAG applications, supporting English, German, Spanish, French, Italian, and Japanese.
Loading preview...
Model Overview
NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-BF16 is a 3-in-1 elastic large language model (LLM) developed by NVIDIA. It contains three nested model variants (30B, 23B, and 12B parameters) within a single BF16 checkpoint, all sharing the same parameter space. These smaller variants can be extracted zero-shot using a provided slicing script, offering significant memory and throughput benefits.
Key Capabilities & Differentiators
- Elastic Architecture: Embeds 12B, 23B, and 30B models in a single checkpoint, reducing memory footprint by 2.14x compared to separate models. This was achieved by post-training the parent Nemotron 3 Nano 30B with only ~160B tokens, a fraction of the original pre-training budget.
- Elastic Budget Control: A novel inference-time mechanism that uses different-sized nested models for thinking and answering phases. The optimal 23B (thinking) -> 30B (answering) configuration achieves up to 16% higher accuracy and 1.9x lower latency compared to single-model approaches.
- Hybrid MoE Architecture: The parent model combines 23 Mamba-2 and MoE layers with 6 Attention layers, featuring 128 experts (6 active per token).
- Strong Reasoning Performance: Benchmarks show the Elastic-30B variant matches or exceeds the parent Nemotron 3 Nano 30B, and the 23B and 12B variants outperform Qwen3-30B-A3B on key reasoning tasks like AIME-2025 and IFBench.
- Quantization Recovery: Demonstrates high accuracy recovery (97-100%) for FP8 and NVFP4 quantized variants.
- Multilingual Support: Supports English, German, Spanish, French, Italian, and Japanese.
Should I use this for my use case?
- AI Agent Systems & Chatbots: Ideal for developers building AI agents, chatbots, or RAG systems that require flexible performance and efficient resource utilization.
- Cost-Sensitive Deployments: The ability to extract smaller, optimized variants and the throughput improvements make it suitable for scenarios where serving costs are critical.
- Reasoning-Intensive Tasks: Excels in reasoning benchmarks, making it a strong candidate for applications demanding high-quality logical inference.
- Prototyping on Consumer Hardware: The 12B and 23B variants in FP8/NVFP4 can fit on lower-tier RTX series GPUs, enabling broader prototyping access.