Nemotron 3.5 Lightning completes tasks 5.36x faster than the most intelligent mo

In Alex Cheema's tests on NVIDIA's DGX Spark, the company's Nemotron 3.5 Lightning finishes agent tasks sooner than Step 3.7 Flash because it generates each token faster; it writes about as much. Cheema says his team worked with NVIDIA to see how Lightning performs locally on the Spark, ran the same 884 agent tasks on 100 models, and compared them on intelligence and speed (thread). Lightning completed tasks 5.36 times as fast as Step 3.7 Flash, which Cheema calls the most intelligent model that runs on the DGX Spark, while retaining 78% of its intelligence (post).

The token counts explain the speedup. How long a task takes depends on how much text the model writes and how fast it writes it. Lightning is much the smaller model but used about the same number of output tokens per task as Step 3.7 Flash, around 2% more on average, so Cheema attributes the 5.36-times speedup to Lightning's much faster per-token decoding on the Spark.

The 78% compares Lightning with Step 3.7 Flash on the team's intelligence measure; it is not presented as a share of tasks solved. The thread does not define that measure or explain how either figure was computed beyond the token comparison. Another post in the same thread shows the opposite case: against Qwen 3.6 35B A3B, Lightning finishes sooner even though Qwen decodes faster, because Qwen writes much more.