Videos Pile Up While You Sleep — MiniMax H3 at Home

Type one line into a chat. Separate prompts get written for each video, they are sent to ComfyUI, and if you simply leave it alone, the videos are waiting when you come back. That is the setup chimolog7 describes (in Japanese) ^{1 2 3}.

According to the creator, you just chat with "Qwen3.8 Q3_S Strata" and it drives MiniMax H3 to mass-produce videos ¹. The creator writes (in Japanese) that if you say "do 5 for me," it writes a separate prompt for each of the 5 in just 6 minutes and sends them to ComfyUI to generate the videos ². And because you can leave it running until it finishes, the creator writes that speed-up know-how for MiniMax became, personally, unnecessary ³.

We are going to take a detour through that speed-up know-how anyway. Every number and impression below is the creators' own report, so it is attributed to them.

A tool for running MiniMax H3 on a GPU you may already own

Can MiniMax H3 even run at home? The entry point is FreeVideo.

The announcement (in English) says video generation now runs right on your laptop ⁴. It needs as little as 8GB of VRAM and 16GB of RAM, with ComfyUI integration, LoRAs and custom workflows, the post says ^{5 6}.

The README points the same way. FreeVideo runs as a ComfyUI plugin ⁷, with launchers for Windows, macOS and Linux ⁸. The minimum is stated as 8 GB of VRAM and 16 GB of RAM ⁹. The license is Apache-2.0 ¹⁰.

There are conditions on the environment. The Linux build targets x86_64 Linux with an NVIDIA GPU and driver 580 or newer ¹¹. On Apple silicon Macs, local generation is offered as a preview ¹².

MiniMax H3 itself has open weights, and ComfyUI's documentation says they let you run the model locally ¹³. Commercial use of locally generated outputs, however, requires a MiniMax commercial license ¹⁴. Check the official page before you decide what the videos are for.

The speed-up that a follow-up test turned inside out

Volume comes down to speed. sep_is_heim writes (in Japanese) that they built the trendy "Jev" into MiniMax H3 to speed up video generation, using it for sparsifying the attention step ¹⁵.

They also note it is fast even though it queries the Jev cloud during generation ¹⁶. As the creator describes it, Jev judges the importance of each layer (the target is 4 steps and 49 layers) and picks a sparsity rate from 1%, 3%, 5% and 10% ¹⁷. On an RTX4070, the creator reports 6 minutes 7 seconds dropping to 3 minutes 34 seconds, a 41.7% reduction ¹⁸.

yonsan434343 quoted that post and wrote up their own test.

The creator writes (in Japanese) that after trying Jev, MiniMax H3 generated 2x faster and they made 255 videos overnight ¹⁹. Under good conditions it reached 2.8x, but averaged out it was a little under 2x, the creator says ²⁰. They also write that they don't notice any degradation so far ²¹.

This is where it gets interesting. When they dug into it, the creator writes, the source of the speed turned out to be sparsification, not Jev: it was just as fast without running Jev at all ²². They add that they are truly grateful anyway, since they found it while trying out that post ²³.

It doesn't stop at knocking down the original. Someone checks with their own hands and writes back exactly what they found. Follow-ups like this are why we keep reading these posts.

Does the face hold when train light sweeps across it?

Fast and plentiful is one thing; quality is the next question. kentdhani tried a scene they call a tough one ²⁴.

The creator writes that they ran MiniMax H3's Ref2V locally in ComfyUI on an RTX5090 ²⁵. Moving light is tough, they say: a passing train throws changing light across a face, which is where identity usually slips ²⁴. The post shows how Ref2V held up there ²⁴. As a way to check quality, it is an easy scene to borrow.

If you want to do this at your own desk

The hardware the creators name above is an RTX4070 and an RTX5090 ^{18 25}. Still, if you want to bring "let it run while you sleep" to your own desk, here is what NVIDIA says about RTX Spark.

• According to NVIDIA's product page, you can generate images and video right on your device, as often as you like, with no subscription and no per-prompt fees [[F101]].
• On the same page, NVIDIA says running locally gives you instant iteration, no API bills, and data that stays on your machine [[F102]].
• According to NVIDIA's blog, the compact desktop configuration puts the same RTX Spark superchip in a small chassis designed for 24/7 operation, as a dedicated local AI system that keeps agents running continuously [[F143]].

On the same blog, NVIDIA says RTX Spark can run models such as Qwen 3.8 Flash Next unmetered and without sending data to the cloud ²⁶. Separately, Strata, the tool in the opening scene, is described in its README as running Qwen3.8-Flash-Next on a normal PC ²⁷. It needs 12 GB of VRAM or more ²⁸.

If you want to use FreeVideo, MiniMax H3 or Strata on an RTX Spark machine, check each tool's official documentation for supported environments first. FreeVideo's Linux build, for example, is for x86_64 ¹¹.

Here are some RTX Spark machines. Check that the software you want to run supports them before you choose.

As the report of 255 videos in one night shows ¹⁹, clips add up fast when you batch them. A separate drive for them keeps things tidy.

In the morning, a screen that shows color and fine detail makes it easier to review what came out.

Read next

For a pipeline that links images, video and music, see images, video and music with local models.

For handing the direction itself over to an AI, continue with the AI motion director examples.