2026 年 10 月五款可本地运行免费 AI 视频生成工具盘点

以下是视频口播的自动转写，可能有错字。

The five best free ai video generators you could run on your own computer. No credits, no per video bill, just your hardware and power licenses vary.

So check them before commercial work. Quick crash course, vram is your graphics card memory offloading moves part of the model into your system ram.

So it fits on a smaller card, it's slower, but it works. Number 5, kandinsky 6.0. This one is new just recently released.

It generates video as synchronized sound and the core is mit licensed. Great for short sound on clips, five second clips around 480p with hd from a separate upscaling stage.

There's an optimized 16 gig setup plus a smaller light model. Number four, onion video 1.5, it's compact for how detailed it looks and great for bringing still images to life.

No generated sound though clips default to five seconds at up to 720p with 1080p from a separate upscaler. Officially, it needs 14 gigs with offloading for lighter setups.

There's a distilled 40p image to video version that runs in as few as eight steps. Number three, when 2.2 really established ecosystem tons of lauras and custom workflows and an apache 2.0 license clips default to about five seconds up to 720p with no built-in sound.

The full a14b model wants 80 gigs or more, but the smaller 5b should fit on an 8 gig card with comfy ui offloading. Number two, ltx 2.5 native audio connected, multi-shot scenes and first and last frame control.

Awesome for fast iterations and keeping a story flowing. The official local example is about five seconds with longer clips, depending on the workflow.

It generates at 960 by 544, then refines up to full hd. The reference setup wants 32 gigs of vram, but quantized versions with offloading go lower and number one, mini max h3, my top pick for quality.

The generations are beautiful and it's great with complex action and visual consistency. Stereo audio image video and audio references plus first and last frame control clips run up to 15 seconds around 768p locally community testers ran comfy's int 8 version on a 16 gig card with offloading making five second 960 by 544 clips with audio a smaller encoder got it to 12 gigs with about 40 gigs of system ram.

Did i miss any? Also, if i get enough interest, i'll make a video on how to install and run them. So let me know in the comments if you want that because if you don't, i'm not going to make it.

It's a lot of work.