The untuned 27B beat the tuned 75B as an agent

I have to admit, a lot of people we're 100% correct to make the suggestion to try this model. I am sorry I ever doubted. The 27B passed every agentic task on a neutral system prompt in 6-9 tool calls. The 75B needed a hand-tuned profile to pass at all and used 2x the turns. For agents, fewer turns beat faster tokens. The two contenders - Nemotron Puzzle-75B-A9B NVFP4, vLLM, PP=2000 across 3 cards, ~65 t/s decode. I made a post about this model. I still think its good for throughput on chatbots and average u