On August 11, 2026, NVIDIA released Nemotron 3.5 Lightning, an open-weight model built for a specific job that frontier models are overqualified for: executing the high-volume, repetitive steps inside long-running AI agents. The pitch is simple. Let the expensive reasoning model make the plan. Let Nemotron 3.5 Lightning do the work.
Specifications and Architecture
Nemotron 3.5 Lightning is a 30 billion parameter Mixture-of-Experts model with only 3 billion active parameters per token. It uses a hybrid architecture combining Mamba-2 state-space layers, sparse attention layers, and multi-token prediction. The context window reaches 1 million tokens, and it supports tool calling, configurable reasoning, and speculative decoding.
The MoE design is central to its efficiency. With 128 routed experts and a shared expert, the model selects six routed experts per token. That means it retains much of the representational capacity of a 30B dense model while doing computation closer to a 3B model's cost. NVIDIA also publishes NVFP4 quantization weights, enabling single-GPU deployment on hardware ranging from DGX Spark GB10 to H100.
Why an Execution-Focused Model Makes Sense
Long-running agents spend most of their time on routine execution rather than difficult reasoning. After a frontier model generates a plan, the agent may perform hundreds of tool calls, file reads, validations, and formatting steps. Using a full-size reasoning model for every action is slow and expensive.
Nemotron 3.5 Lightning is positioned as the execution layer in a system of models. Complex reasoning and orchestration still demand frontier capability. High-volume, domain-specific tasks are better handled by a fast, accurate, customizable model that can be post-trained for specialized workflows.
Benchmark Performance
NVIDIA published results across knowledge, reasoning, coding, agents, instruction following, and long-context tasks. Selected scores:
- MMLU Pro: 81.94 (BF16), 81.62 (NVFP4)
- SWE-bench Verified: 51.56 (BF16), 52.80 (NVFP4)
- BrowseComp: 36.97 (BF16), 36.81 (NVFP4)
- PinchBench: 85.37 (BF16), 83.43 (NVFP4)
The notable pattern is how little NVFP4 quantization degrades performance. On SWE-bench Verified, the quantized model actually scores higher than BF16, likely within run variance. This matters for deployment, because it means enterprises can run the model at lower precision without significant quality loss.
Availability and Licensing
The model is available under the OpenMDW 1.1 license through NVIDIA's API and on partner platforms like FriendliAI. It is intended primarily for English and programming languages, with Spanish, French, German, Italian, and Japanese also supported.
NVIDIA is positioning Nemotron 3.5 Lightning as a workhorse for always-on agents: PR summarization and test triage in software development, document extraction in financial services, alert enrichment in security operations, and long-running personal agents for email and calendar. The common thread is high call volume on well-scoped tasks, exactly where inference efficiency compounds with every step.