NVIDIA Nemotron 3.5 Lightning 30B A3B

NVIDIA Nemotron 3.5 Lightning 30B A3B, released August 1, 2026, is an efficient open-weight model designed for local and single-GPU deployment. Its 30B total parameters activate only 3B per token, using a hybrid Mamba-2 + MoE + attention architecture that balances throughput with strong conversational and coding ability. With up to 1M-token context and a configurable reasoning mode, it targets developers who want frontier-adjacent capability on hardware they already own.

Model Specifications

Key Features

• 3B active parameters: A 30B-class model with the throughput of a much smaller model, fitting a single 80GB GPU.
• Configurable reasoning: Thinking mode toggles via the chat template (enable_thinking=True/False), letting you switch between fast responses and deep reasoning.
• DSpark speculative decoding: Optimized for low-concurrency datacenter deployments to cut latency.
• Long context: Up to 1M tokens for retrieval-heavy and agentic workloads; 256K is the practical default on one H100.
• Built to customize: Best-for-use post-training (SFT, RL, distillation), domain adaptation, and building quantized variants.

Use Cases

• Local coding assistants: Run a capable reasoning model on a single data-center GPU without API dependence.
• Model customization: Fine-tune or distill the full-precision weights for domain-specific tasks.
• Research and evaluation: The BF16 reference weights make it a clean baseline for studying efficient MoE hybrids.

Advantages

1. Efficiency: 10x sparsity (30B total, 3B active) gives a favorable speed-to-quality ratio on both Hopper and Blackwell.
2. Permissive model: OpenMDW-1.1 licensing plus released post-training datasets support open research and commercial customization.
3. NVIDIA ecosystem: Native integration with NVIDIA NIM, build.nvidia.com, and the wider Nemotron toolchain.

Getting Started

Chat with the model at build.nvidia.com/nvidia/nemotron-3.5-lightning-30b-a3b or download the BF16 weights from Hugging Face (nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16). For local use, load it in transformers or vLLM on an 80GB GPU, enable thinking for hard tasks, and consider DSpark for latency-sensitive serving.

Alternatives

• Muse Glimmer: Meta's 30B dense agentic model that fits 24GB consumer cards.
• Qwen 3.6 27B: Dense competitor with strong benchmark performance at a similar size.
• DeepSeek V4: Larger open MoE option with different efficiency tradeoffs.

Conclusion

Nemotron 3.5 Lightning 30B A3B is a strong pick when you need an open, efficient, single-GPU reasoning model with long context and deep customization hooks. Its hybrid architecture and 3B-active design make it one of the most practical frontier-adjacent options for on-premise deployment today.