NVIDIA: Nemotron 3.5 Lightning (free)

NVIDIA Nemotron 3.5 Lightning is an open mixture-of-experts language model designed to balance large total capacity with efficient inference. It carries roughly 30 billion total parameters while activating only about 3 billion per token, a sparse design that keeps compute costs low without giving up the reasoning depth expected from much larger dense models. This active-parameter profile makes the model especially attractive for high-throughput agentic workflows where many parallel requests need to be served quickly. Because the weights are publicly released, teams can fine-tune or distill the model for specialized enterprise domains, adapting it to proprietary data without depending on a closed vendor stack.

In practical deployment, Nemotron 3.5 Lightning is positioned for agentic tasks, tool-driven pipelines, and structured-output scenarios that benefit from custom post-training. The architecture's sparsity, combined with open-weight availability, allows organizations to run it on their own infrastructure or route it through gateways that aggregate multiple NVIDIA-hosted endpoints. Its emphasis on reliability across enterprise workloads suggests it is well suited for production assistants that must handle consistent reasoning, controlled generation parameters, and integration with external tools rather than purely open-ended chat. For builders looking for a capable, openly licensed model that can be specialized, this release fits neatly between smaller on-device models and much larger frontier systems.