Field note
Kilo added NVIDIA Nemotron 3.5 Lightning to its CLI and VS Code products and evaluated it inside the company's coding-agent harness. The model is described as a 30-billion-parameter Mixture-of-Experts system with three billion active parameters per token and context up to one million tokens.
Kilo reports 100 percent on a deterministic markup task for the Instant variant and 73 percent exact match on its personal git evaluation for Thinking. These are vendor results on Kilo-selected tasks, not an independent general-purpose model ranking.
The useful framing is a constrained executor. A relatively small active footprint can be attractive for frequent tool-driven steps if the harness narrows the task and verifies outputs. Teams considering it should replay their own repository tasks and measure corrections, latency, and total cost rather than transferring headline scores directly.