XNews

NVIDIA Nemotron 3.5 Lightning Model Release

NVIDIA Nemotron 3.5 Lightning Model Release

NVIDIA released Nemotron 3.5 Lightning, an open-weights Mixture of Experts model with 30 billion total parameters and 3 billion active parameters, supporting 1 million token context length. The model offers local deployment options in BF16 and smaller NVFP4 formats, running on a single H100 GPU or DGX Spark, with RTX 5090 also supported. Official figures claim up to 4x faster output speeds and 86% accuracy on PinchBench, 30% faster than Qwen3.6 35B at comparable accuracy. The model is available through the Perplexity Agent API at $0.0115 per 1M input tokens and $0.17 per 1M output tokens. Perplexity CEO Aravind Srinivas promoted it as an American open-weights MoE model running efficiently on laptops and local hardware like DGX Spark, with larger Nemotron Ultra also available on Perplexity.

Key insights

Nemotron 3.5 Lightning Delivers Frontier Performance in a Locally Deployable Package

The 30B total with only 3B active parameters enables the model to run on consumer-grade hardware including single H100, DGX Spark, and RTX 5090. This architecture achieves 86% PinchBench accuracy while being 30% faster than Qwen3.6 35B at equivalent accuracy. Output speed improvements of up to 4x over previous versions make real-time applications feasible on local hardware.

Perplexity API Integration Provides Immediate Developer Access with Competitive Pricing

Perplexity Agent API pricing at $0.0115 per 1M input tokens and $0.17 per 1M output tokens positions Nemotron 3.5 Lightning competitively for developer adoption. Availability through Perplexity alongside the larger Nemotron Ultra model creates an upgrade path. The 1M token context window supports long-document processing and extended agent interactions.

Source posts