Can a Single DGX Spark Generate 10 Seconds of AI Video in Under 5 Minutes?

MiniMax H3 just got a LOT more practical to run locally. In my previous video, I showed my first results after finally getting MiniMax H3 working on my NVIDIA DGX Spark. Since then, I discovered the MiniMax H3 4-Step Turbo LoRA, and I had to test it. The difference in generation time was huge. In my testing, the 4-step workflow cut generation times roughly in half—and sometimes by even more. My measured results on a single DGX Spark: ⚡ 10-second videos: ~4 minutes 46 seconds average ⚡ 15-second video: ~7 minutes 12 seconds ⚡ Only 4 sampling steps That speed completely changed how I used H3. Instead of generating isolated test clips, I started experimenting with: • Image-to-video • Character consistency • Multi-scene continuations • Dialogue and lip sync • Music and sound effects • Cinematic camera movement • Different animation styles • Fight scenes • Longer AI-generated stories Eventually, those experiments turned into an entire animated detective-vampire story, which I also show in this video. Everything shown from MiniMax H3—including the video, dialogue, music, ambience and sound effects—was generated locally on my NVIDIA DGX Spark. The Turbo LoRA used in this test is larryvrh/MiniMax-H3-Turbo-Lora on Hugging Face. Subscribe for more local AI testing, benchmarks, models, video generation and DGX Spark experiments. #MiniMaxH3 #DGXSpark #LocalAI #AIVideo #AIGeneration #ComfyUI #NVIDIA #GenerativeAI #AIAnimation #OpenSourceAI #AITools #MiniMax