Prime Intellect and NVIDIA lift Nemotron's enterprise score by 70% with RL
A 2,500-task training set raised the model's score on AutomationBench-AA's public split from 34.0% to 57.7% in about 11 hours.
By Ryan Merket · Published · Updated
Primary source: X
Why it matters
The collaboration makes Prime Intellect's enterprise-agent pitch tangible: build workflow tasks with checkable outcomes, use them to post-train an open model, then evaluate the result. The public-split gain is meaningful, but the score does not mean that 57.7% of tasks were fully completed or prove performance in live company systems.
Prime Intellect, co-founded by Vincent Weisser (@vincentweisser) and Johannes Hagemann (@johannes_hage), worked with NVIDIA to post-train Nemotron 3.5 Super VL for enterprise workflows, reporting a 70% relative gain on AutomationBench-AA's public split. The work, published October 15th, took the model from 34.0% to 57.7% after NVIDIA ran reinforcement learning on 2,500 tasks built with Prime Intellect.…