NVIDIA: Nemotron 3.5 Lightning

NVIDIA: Nemotron 3.5 Lightning is an open mixture-of-experts text model from NVIDIA. It has 3B active parameters out of 30B total, a design that keeps only part of the network active per token. The provider released it on 2026-08-11. It sits in the text category, so it is built for language work rather than image or audio input.

The model accepts text and returns text. Its context window is 262,144 tokens, and the maximum output is 235,929 tokens, so long inputs and long generations are both possible within those bounds. Because it is a mixture-of-experts model with a small active parameter count, NVIDIA positions it for high-throughput agentic workloads and specialized tasks. It also supports reasoning output, which you can control through the reasoning and include_reasoning parameters.

Those capabilities point to a few practical uses. Long context plus large output makes it a fit for processing lengthy documents, such as contracts, logs or research material, and producing detailed summaries or extractions. Tool calling support makes it suitable for agent loops where the model decides which function to invoke and feeds results back into the conversation. Structured outputs and response_format help when you need JSON or another schema-conforming shape for downstream code. The reasoning controls let you enable or inspect intermediate reasoning when a task needs it, and turn it off when latency or output size matters more.

You call it through api14's OpenAI-compatible API using the model id "nvidia/nemotron-3.5-lightning". The supported request parameters are frequency_penalty, include_reasoning, logit_bias, logprobs, max_tokens, min_p, presence_penalty, reasoning, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs and top_p. Sampling can be tuned with temperature, top_p, top_k and min_p, while seed, stop and the penalty parameters give you reproducibility and control over repetition. Tool use is configured with tools and tool_choice, and output shape with response_format and structured_outputs.

Keep the limits in mind when choosing it. Input is text only, so anything involving images or audio needs a different model. The context window and maximum output are large but finite, and requests beyond them will not fit. If your workload does not need tool calling, structured output or reasoning, a simpler text model may be enough.