NVIDIA released Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model with 3 billion active parameters at inference time and a 1 million token context window, alongside NeMo Switchyard, an open-source model routing library, on August 11, 2026. External evaluations published since the launch have added both supporting data and qualifications to NVIDIA's claims.
On Nemotron 3.5 Lightning: Artificial Analysis rated it at 24 on their Intelligence Index, above the median of 9 for comparable open-weight models, and measured output speed at 296.7 tokens per second against a median of 101.6 t/s for the class. However, a review found that the model's 1 million token context window does not produce best-in-class long-context comprehension: on NVIDIA's own long-context benchmark, it scored 52.0, behind Qwen3.6 35B (61.06) and Gemma 4 26B (57.56). Competitive analysis places Lightning behind open models Qwen3.6 35B A3B and Muse Glimmer and behind closed models such as Gemini 3.6 Flash on general agent and coding evaluations, though the model showed major benchmark gains over its predecessor Nemotron 3 Nano.
On NeMo Switchyard: a detailed Cognition case study showed the router operating between Opus 5 and Kimi K2.7 on a production coding benchmark, achieving 50.6% accuracy at a $3.11 mean cost, which Cognition reported as approximately 28% lower than using Opus 5 alone. A separate review of version 0.2 concluded the library is not ready as a drop-in production dependency, citing documented gaps including upstream work continuing after client disconnect, undercounted retry recovery, and missing version headers.