How does it compare to Qwen 3.6 35B A3B, one of the most popular models on DGX Spark today?
Qwen with MTP actually runs at faster decode TPS than Lightning for most context lengths and tasks.
However, Qwen is much more verbose: it uses 45% more output tokens and 11% more steps to complete tasks. This histogram shows the output token distribution over all GDPVal tasks.
Overall tasks take 42% longer with Qwen than Lightning.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
Why it matters
Reframes model selection around tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition →-to-completion rather than tokens-per-second, which is the number that actually sets AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → latency and cost.
