/tags/nvidia/ Recent content in Nvidia on AI Prompt Toolkit Hugo en-us Fri, 12 Jun 2026 00:00:00 +0000 /prompts/cosmos3-super-master-prompts/ Fri, 12 Jun 2026 00:00:00 +0000 /prompts/cosmos3-super-master-prompts/ <h2 id="cosmos3-super-prompt-guide">Cosmos3-Super Prompt Guide</h2> <p><strong>NVIDIA Cosmos3-Super</strong> (released June 2026) is a <strong>64 billion parameter physical-AI omnimodel</strong> — a world-model architecture that combines a 32B reasoning module with a 32B generation module. Unlike traditional video generators that predict pixels, Cosmos3-Super <strong>simulates physics</strong> and couples action trajectories with synchronized video and audio output.</p> <h3 id="architecture">Architecture</h3> <pre tabindex="0"><code>Action Trajectory → [32B Reasoner] → Physical State → [32B Generator] → Video + Audio ↑ World Knowledge </code></pre><h3 id="prompting-strategy">Prompting Strategy</h3> <p>Cosmos3-Super requires a fundamentally different prompting approach than diffusion-based video models (Sora, Runway, Kling):</p> /prompts/nemotron-3-ultra-master-prompts/ Fri, 12 Jun 2026 00:00:00 +0000 /prompts/nemotron-3-ultra-master-prompts/ <h2 id="nemotron-3-ultra-prompt-guide">Nemotron 3 Ultra Prompt Guide</h2> <p><strong>NVIDIA Nemotron 3 Ultra</strong> (released June 2026) is the first open-weight <strong>550 billion parameter hybrid Mamba–Mixture-of-Experts model</strong> — a groundbreaking architecture combining Mamba&rsquo;s linear-time sequence processing with Transformer-based expert modules.</p> <h3 id="architecture">Architecture</h3> <pre tabindex="0"><code>Input → [Mamba Backbone] → [MoE Router] → [Expert 1..N] → Output ↑ Linear time ↑ 55B active ↑ Sparse activation 1M context OK out of 550B total ~10% active params </code></pre><h3 id="key-specifications">Key Specifications</h3> <table> <thead> <tr> <th>Metric</th> <th>Value</th> </tr> </thead> <tbody> <tr> <td>Total Parameters</td> <td>550B</td> </tr> <tr> <td>Active Parameters</td> <td>55B (~10%)</td> </tr> <tr> <td>Context Window</td> <td>1,000,000 tokens</td> </tr> <tr> <td>MMLU Score</td> <td>89.1</td> </tr> <tr> <td>Architecture</td> <td>Hybrid Mamba–Transformer MoE</td> </tr> <tr> <td>License</td> <td>Open weights (NVFP4 variant on Hugging Face)</td> </tr> </tbody> </table> <h3 id="prompting-strategy">Prompting Strategy</h3> <p>Nemotron 3 Ultra&rsquo;s unique Mamba-MoE architecture requires different prompting than pure Transformer models:</p>