NVIDIA Nemotron 3.5 Lightning

Overview

Nemotron 3.5 Lightning is an open-weight hybrid model from NVIDIA with 30 billion total parameters and roughly 3 billion activated per token. It chains together Mamba-2, attention and mixture-of-experts layers, was pretrained on more than 20 trillion tokens, and pairs an NVFP4 recipe with multi-token prediction for faster generation. Context reaches up to 1M tokens and the model supports both reasoning and tool calling, covering English, coding languages, plus Spanish, French, German, Italian and Japanese. Weights are released under the OpenMDW License Agreement v1.1.

Its positioning is not standalone. NVIDIA pairs it with Nemotron 3 Ultra: Ultra is the heavy model at 550 billion total and 55 billion active parameters for tasks needing long reasoning chains, while Lightning takes on tool calling, coding and structured output, the high-frequency agent execution steps with clear boundaries, and works well as a resident execution unit for sub-agents.

The BF16 benchmarks NVIDIA publishes are MMLU Pro 81.94, GPQA Diamond 75.44, SWE-bench Verified 51.56, IFBench 71.88, AA-LCR 52.00 and Terminal-Bench 2.1 at 24.58, all on the vendor side of the ledger. Hosted API list prices sit around 0.07 to 0.08 dollars per million input tokens and 0.20 dollars for output, and OpenRouter published an explainer on September 22 covering its role in high-frequency agent execution calls.

Local deployment is the other selling point. A third-party test measured 334 tokens per second on an RTX 5090 with a first token after a 128k document at 14.9 seconds; on an RTX 3090 running a quantized GGUF build it reached 193.7 tokens per second, and moving context from 32k to 192k added only 1,119 MiB of VRAM. That memory efficiency at long context is what lets it run on consumer cards.

Key Features

• Mamba-2 hybrid architecture: Mamba-2, attention and mixture-of-experts layers chained together, cutting per-token work and holding long sequences in less memory.
• 30B total, 3B active: Only about 3 billion parameters wake up per token, keeping speed and cost in the lightweight tier.
• Up to 1M-token context: Long documents and long tool traces can be processed whole, with hosted endpoints commonly configured at 262K.
• Reasoning and tool calling: Supports reasoning mode and function calling, suited to serving as an agent execution unit or resident sub-agent node.
• Open weights under OpenMDW: Released under the OpenMDW License Agreement v1.1, so it can be quantized and deployed locally.
• Multilingual coverage: Beyond English and coding languages, covers Spanish, French, German, Italian and Japanese.

Use Cases

• High-frequency agent execution steps such as tool calling, structured output and format conversion
• Sub-agent execution units that handle the low-cost resident role inside long agent chains
• Local and edge deployment of quantized builds for long-document question answering and summarization
• Multilingual batch processing beyond English, including Spanish, French, German, Italian and Japanese content

Pros

• 3B active parameters on a Mamba-2 hybrid architecture give fast generation and small memory growth at long context
• Open weights under OpenMDW can be quantized and run on consumer graphics cards
• Context reaches 1M tokens, so long documents and tool traces need no splitting
• Pairs with Nemotron 3 Ultra, letting one stack allocate work by task difficulty

Pricing

Hosted API list prices run about 0.07 to 0.08 dollars per million input tokens and 0.20 dollars for output, with some gateways offering it on a free tier. Weights are open under the OpenMDW License Agreement v1.1 for self-hosting. Check individual providers for current rates.

Summary

Nemotron 3.5 Lightning is an open-weight hybrid model from NVIDIA at 30 billion total and roughly 3 billion active parameters, combining Mamba-2, attention and mixture-of-experts layers, pretrained on more than 20 trillion tokens, with up to a 1M-token context, reasoning and tool calling, and weights under the OpenMDW license. NVIDIA pairs it with the 550B Nemotron 3 Ultra so it can take on high-frequency execution steps like tool calling and coding. Hosted API pricing sits around 0.07 to 0.08 dollars per million input tokens, and third-party consumer-card tests measured several hundred tokens per second with only a small memory increase at long context.

Version History

• OpenRouter explains the role of Nemotron 3.5 Lightning in high-frequency agent execution (2026-09-22): OpenRouter published an explainer on NVIDIA Nemotron 3.5 Lightning, a 30B total, roughly 3B active mixture-of-experts open-weight model positioned for tool calling, coding and other high-frequency agent execution steps with clear boundaries, paired in a division of labour with Nemotron 3 Ultra (550B total, 55B active) for complex reasoning. The model combines Mamba-2 with attention and mixture-of-experts layers, was pretrained on more than 20 trillion tokens, supports up to a 1M-token context, and ships under the OpenMDW License Agreement v1.1.
• NVIDIA open-sources Nemotron 3.5 Lightning, a 30B-A3B hybrid (2026-08-11): NVIDIA released and open-sourced Nemotron 3.5 Lightning, a hybrid model with 30B total and roughly 3B active parameters combining Mamba-2, attention and mixture-of-experts layers, with an NVFP4 recipe and multi-token prediction, pretrained on more than 20 trillion tokens, up to a 1M-token context, supporting reasoning and tool calling, and covering English, coding languages plus Spanish, French, German, Italian and Japanese under the OpenMDW License Agreement v1.1. NVIDIA-published BF16 benchmarks are MMLU Pro 81.94, GPQA Diamond 75.44, SWE-bench Verified 51.56, IFBench 71.88, AA-LCR 52.00 and Terminal-Bench 2.1 at 24.58.