NVIDIA AI-Q Tops DeepResearch Benches I & II

🔑 Enhanced Key Takeaways

• •NVIDIA AI-Q is powered by the newly released Nemotron 3 Super model, a 120-billion-parameter open-source system launched on March 11, 2026.[1][3]
• •Nemotron 3 Super employs a hybrid mixture-of-experts architecture with Mamba and transformer layers, activating only 12 billion parameters during inference for 5x higher throughput.[1][3]
• •DeepResearch Bench consists of 100 PhD-level tasks across 22 fields, testing multistep research on large document sets while maintaining reasoning coherence.[1][6]

🛠️ Technical Deep Dive

• •Hybrid MoE architecture: Combines Mamba layers (4x higher memory/compute efficiency) with transformer layers for reasoning; only 12B of 120B parameters active at inference.[1]
• •Latent MoE: Activates four expert specialists for the cost of one token generation, improving accuracy.[1]
• •Multi-Token Prediction: Generates multiple future words simultaneously for 3x faster inference.[1]
• •1-million-token context window to retain full workflow state and prevent goal drift.[1]
• •Optimized for Blackwell platform in NVFP4 precision: 4x faster inference than FP8 on Hopper with no accuracy loss.[1]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline