Many AI models are huge, which makes them slow and expensive to run. Engineers need models that can reason, write code, and see images, but that can also respond quickly, especially when a program is making many tool calls. This paper creates two efficient models.
Nemotron 3.5 Super uses a smart design where only a fraction of the model activates for any given input. It was trained on a massive amount of software engineering data to improve its coding ability. The model was also fine-tuned with feedback from verifiers and from human preferences, and a technique called multi-teacher was used to combine skills from specialized teachers.
The smaller model beat a much larger competitor on a coding benchmark while using roughly 4.6 times fewer active parameters. It also achieves fast output rates, like 4,856 tokens per second per GPU on a certain workload. This shows that careful training can deliver strong performance without requiring huge models.