I believe the future of AI belongs to local deployment rather than cloud-based models. First, the gap between open-source and closed-source models has narrowed significantly. In many cases, open source is the first-choice option—it can handle most general agent tasks. Second, advances in the underlying algorithms, including improvements to inference engines, make it possible to run models locally (and I mean in production environments). Compute boxes like the NVIDIA DGX Spark further accelerate this trend. I believe more companies will build similar boxes. And don't forget that Apple has a huge number of users on laptops, PCs, and desktop boxes like the Mac mini and Mac Studio

First attempt at building a product ad video for tesla cybertruck with ultraxvideo. Not perfect, but the proof is clear: agent-driven video production works — and this beats what most people are doing, efficiently. Full 1080p, all rendered on an RTX 3090.

Agent 比人做的更好，而且可以自动化执行，视频在 RTX 5070 12G vram 上完成，底层模型 minimax h3, music 3。字幕、片头、音乐设计、视频设计都通过 ultraxvideo 完成。seedvr2 超分到 720p, video2x 继续超分到 1080p seedvr2在 12G vram 实际可超分到 1080p.

So goood to have thus kind of 1080p video on 12G vram with MiniMax H3. 10s => 44 minutes, 640p upscale to 1080p w/ SeedVR2

Now I can generate 10-second 640p videos locally and upscale them to 1080p. P.S. 480p videos of 15–20 seconds work fine already. MiniMax H3 RTX 5070 12G

这是 H3 生成的原视频，640p upscale 到了 1080p

我一直想，既然MiniMax H3在视频生成上做得这样好，它应该也适用于图像理解和生成。基于这点，我用 image prompt 创建了几个 1s/2s 的短视频，我想这的确证实了我的猜想，这些图片截自H3 生成的视频

Here is another example, 720p, all done on just 12GB VRAM. MiniMax H3 will empower AI video creation.