Nemotron 3.5 Lightning

// Overview

Nemotron 3.5 Lightning is the first model in NVIDIA's Nemotron 3.5 generation, published on 11 August 2026 alongside NeMo Switchyard, an open-source library that routes each step of an agent workflow to the model that fits it. NVIDIA positions Lightning not as a frontier reasoner but as the high-volume execution layer underneath one: the model that handles the many cheap, repetitive steps a long-running agent takes between its occasional hard decisions.

The architecture is a hybrid — Mamba-2 state-space layers interleaved with attention layers on top of a Mixture-of-Experts feed-forward stack — with 30 billion total parameters and 3 billion active per token. It carries the same shape and small size as Nemotron 3 Nano while scoring substantially higher across NVIDIA's evaluation suite. The context window runs to one million tokens, though NVIDIA's own serving guidance caps a single H100 at 256K because of memory rather than the model.

Weights ship under the permissive OpenMDW-1.1 license in two checkpoints — a BF16 release aimed at customisation and research, and an NVFP4 quantisation NVIDIA recommends for production inference — together with the pre-training and post-training datasets and the recipes used to build them. The model is served through vLLM, SGLang and TensorRT-LLM, supports speculative decoding via DSpark, DFlash or MTP, and is distributed through Hugging Face, ModelScope, OpenRouter and NVIDIA's own NIM microservices on build.nvidia.com. Supported languages are English, Spanish, French, German, Italian and Japanese, plus code.

NVIDIA's published table sets Lightning against Qwen 3.6 35B A3B, Gemma 4 26B A4B, GPT-OSS 20B and its own Nemotron 3 Nano and Super. It reads honestly: Lightning scores 81.94 on MMLU Pro, 75.44 on GPQA Diamond and 51.56 on SWE-bench Verified, well clear of Nemotron 3 Nano on every row but behind Qwen 3.6 35B A3B on most of them — including SWE-bench Verified (70.12), Terminal-Bench 2.1 (44.38) and GPQA Diamond (83.40). Its case is speed rather than peak accuracy: NVIDIA reports up to 4× the output speed of similarly-sized models, and its developer blog claims 10,000 PinchBench tasks completed 30% faster than Qwen3.6 35B at comparable accuracy.

// Benchmarks

This model's scores

1. PinchBench85.37%
2. MMLU Pro81.94%
3. GPQA Diamond75.44%
4. IFBench71.88%
5. AA-LCR52%
6. SWE-bench Verified51.56%
7. SWE-bench Multilingual39.33%
8. BrowseComp36.97%
9. SciCode32.6%
10. Terminal-Bench 2.124.58%
11. AA-Omniscience17.5%
12. Humanity's Last Exam11.72%
13. τ³-bench (Banking)9.28%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

OpenRouter's hosted rate for the open-weight model at 262K context — NVIDIA publishes weights rather than a first-party token price. A free 1M-context endpoint is also listed.

Pricing source ↗

// Strengths

• Permissively licensed under OpenMDW-1.1, with weights, pre-training and post-training datasets and recipes all published
• 3B active parameters out of 30B keeps per-token compute — and cost — low for agent loops that make thousands of calls
• 1M-token context, with NVIDIA-documented serving configurations for vLLM, SGLang and TensorRT-LLM
• Speculative decoding supported out of the box through DSpark, DFlash or MTP draft strategies
• 81.94 on MMLU Pro and 85.37 on PinchBench on NVIDIA's own evaluations — ahead of GPT-OSS 20B and Nemotron 3 Nano across the board

// Best for

• The execution tier of an agent stack, handling routine tool calls and sub-steps while a larger model makes the hard decisions
• Self-hosted deployments where a permissive license and published training data matter more than topping a leaderboard
• High-throughput batch work — classification, extraction, summarisation — where cost per token dominates
• Research and customisation on an open hybrid Mamba-2/MoE architecture, using the BF16 checkpoint and NVIDIA's released recipes

// How to access