Nvidia’s Nemotron 3.5 Lightning is designed for the many small model calls that happen after an AI agent has made its plan. It can choose tools, generate arguments, edit files and validate intermediate results without invoking a larger reasoning model for every step.
The open-weight model contains 30 billion parameters but activates roughly 3 billion for each token. This mixture-of-experts design reduces computation while retaining more total capacity than a dense 3-billion-parameter model. The full checkpoint must still fit in storage and memory, so the active count is not a complete measure of hardware needs.
Nvidia combines Mamba-2, attention and expert layers, and provides BF16, NVFP4 and GGUF weights under its OpenMDW 1.1 license. The model card specifies a context window up to one million tokens. OpenRouter’s standard endpoints currently expose 262,144 tokens, while its free endpoint lists the full million; completion limits and structured-output support vary by provider.
Lightning is meant for well-scoped execution rather than the hardest planning calls. Developers should confirm current endpoint limits and test tool-use reliability, because lower latency is useful only if repeated errors do not force expensive retries.