Training small orchestration agents to solve big problems is a powerful new approach in AI system design. At NVIDIA Research, we’ve developed ToolOrchestra, a method that uses a small, specialized model—called an orchestrator—to manage and coordinate larger models and tools in solving complex tasks. The orchestrator acts as a decision-making supervisor, evaluating the task, user preferences, and available resources to choose the right model and tool at each step, balancing speed, cost, and accuracy. What makes this approach effective is that small models, when properly trained, are well-suited for orchestration. They are not burdened by vast knowledge, which can lead to overthinking or irrelevant responses. Instead, their limited size forces them to focus on the core logic of problem-solving, making them more efficient and precise in decision-making. The key to ToolOrchestra’s success lies in its training process. It uses synthetic data generation to create thousands of task-solving scenarios, then applies multi-objective reinforcement learning. This training explicitly rewards high accuracy, low cost, and fast execution—aligning the orchestrator with real-world constraints. The result is a model that doesn’t just follow instructions, but makes smart, cost-aware decisions. In tests, Orchestrator-8B, a small 8-billion-parameter model, outperformed much larger models—including GPT-5, Claude Opus, and Llama-3.3-70B—on difficult benchmarks like Humanity’s Last Exam, FRAMES, and τ2-Bench. It achieved higher accuracy while using significantly less time and cost. Even when limited to just 10 or 20 reasoning turns, Orchestrator-8B maintained its lead, showing that it can deliver high performance under tight constraints. The training process is surprisingly lightweight. Using just 552 synthetic problems and 1,296 training prompts, and starting with a small model like Qwen3-8B, the method produces a high-performing orchestrator. The steps are simple: pick a base model, generate synthetic data, train with reinforcement learning, and monitor progress using tools like wandb. This approach transforms the way we build AI agents. Instead of relying on large, monolithic models to do everything, we use small, efficient orchestrators to direct them. This reduces cost, improves speed, and makes systems more adaptable. It’s a shift from brute-force scaling to smart, strategic coordination. Looking ahead, this work points to a new era of compound AI systems—where multiple specialized models and tools work together under intelligent control. These systems are not only more powerful than single large models but also safer, faster, and more sustainable. ToolOrchestra is a foundational step toward this future, proving that small models, when orchestrated well, can solve big problems better than ever before.