Prime Intellect Lab adds Nemotron 3 Ultra

Turnkey post-training platforms are democratizing the creation of custom AI agents, making the raw intelligence of base models less of a moat than a company's proprietary workflow environments and RL loops.

• –Turnkey Infrastructure: By offering pre-configured environments, sandboxes, and verifiers, Lab removes the substantial engineering overhead that historically prevented enterprises from training their own RL models.
• –Hardware and Serving Optimization: Running and fine-tuning a 550B MoE model is cost-prohibitive for most teams; native Blackwell support and NVIDIA Dynamo inference integration make optimization financially and operationally viable.
• –Continuous Iteration: The platform's ability to feed real-world production traces back into training runs solves the static evaluation problem, enabling agents to improve dynamically over time.
• –Specialization Outperforms Scale: As proven by case studies like Ramp, smaller, task-specialized open models fine-tuned with RL can outperform much larger generalist models like Claude Opus at lower latency and cost.