别让米其林主厨削土豆，英伟达用“小脑指挥大脑”，重构AGI生产力

Core Insights - NVIDIA has introduced the Orchestrator model with 8 billion parameters, which significantly reduces costs and improves efficiency compared to larger models like GPT-5, achieving a score of 37.1% on the HLE benchmark while costing only 30% of GPT-5's expenses [1][16]. Performance and Cost Efficiency - Orchestrator outperforms GPT-5 in multiple benchmarks, achieving 80.2% accuracy on τ2-Bench and 76.3% on FRAMES, while also reducing inference costs to 9.2 cents per task, which is 30% of GPT-5's cost [16][20]. - The model demonstrates a strong ability to generalize to unseen tools and maintains performance with minimal fluctuations when faced with new pricing strategies or models [22]. Model Architecture and Training - The Orchestrator employs a unique architecture that separates decision-making from execution, utilizing a lightweight scheduling model to optimize task allocation among various specialized tools [6][23]. - The training process incorporates a reinforcement learning framework that balances accuracy, efficiency, and user preferences, leading to a model that is both cost-effective and adaptable [10][11]. Innovation in AI Systems - The introduction of Orchestrator signifies a shift towards composite AI systems that leverage multiple models and tools, offering advantages in safety, speed, and cost over traditional single large models [23]. - This approach marks a potential new paradigm in AI development, moving away from reliance on a single powerful model to a more efficient and scalable system [23].