NVIDIA Releases Nemotron 3.5 Lightning for High-Volume Agent Work

NVIDIA has released Nemotron 3.5 Lightning, an open-weight model designed to handle the repetitive execution steps inside long-running AI agents. According to LLM Stats, the model launched on August 11, 2026, with 30 billion total parameters but roughly 3 billion active for each task.

The hybrid architecture combines Mamba-2, mixture-of-experts routing, and attention. NVIDIA positions it as an execution model rather than a replacement for a frontier reasoning system: one model can create the plan, while Lightning handles tool calls, command output, file operations, and other frequent lower-level steps. Its NeMo Switchyard companion is intended to automate that division of labor.

Built for local and hosted workflows

Nemotron 3.5 Lightning supports a context window of up to 1 million tokens, although the Hugging Face documentation reportedly puts practical single-H100 usage at about 256,000 tokens. It is available in BF16 for customization and NVFP4 for deployment, with weights, data, and recipes released under NVIDIA’s OpenMDW-1.1 license. The model can be accessed through NVIDIA’s platform, OpenRouter, and other inference providers; hosted pricing was not yet listed.

NVIDIA claims output speeds of up to four times those of similarly sized models. On PinchBench, it reports an 86% score and performance about 30% faster than Qwen3.6 35B at comparable accuracy. However, its reported Terminal-Bench 2.1 result is approximately 24.6, suggesting builders should treat it as a fast executor—not necessarily their primary coding agent.

The release also includes speculative-decoding support through built-in multi-token prediction, plus DSpark and DFlash helpers. Compatibility spans tools including vLLM, SGLang, TensorRT-LLM, llama.cpp, Ollama, and LM Studio, making it relevant for teams comparing cloud latency with local inference costs.

Source: LLM Stats