This Smaller DeepSeek Model Scored 53% Higher DeepSeek-V4-Flash-0731 scored 82.7 on Agentic Coding Eval.
Nemotron-3-Ultra scored 53.9.
But here's what makes it interesting:
284B vs. 550B total parameters. 13B vs. 55B active parameters.
DeepSeek is using far fewer active parameters while scoring significantly higher. The model isn't just smaller.
It's more efficient.