We queried Hugging Face this morning while putting together the daily note. The MiniMax H3 repository was created on 28 July, and it currently shows about 5.5 million downloads.

Then we counted how many other repositories carry its name. Roughly 400. New ones are landing at between eight and twenty-four per day, every day, for the last ten days straight.

One honest caveat before anyone quotes that figure. We counted repository names, not achievements. That total includes straight mirrors, re-uploads, abandoned experiments and a fair amount of noise, so it is a measure of activity rather than of four hundred distinct good ideas. It is still a remarkable amount of activity.

The work changed shape this week

What makes today interesting is not the count. It is what the new repositories are for.

Two days ago the community was solving whether H3 could run at all on normal hardware. That produced a chunked sampler for long video on 16 GB, a node that trades quality for speed, and a fork that splits the work across two consumer cards. We wrote that up yesterday.

Today the same people have moved on to making it fast and making things with it. The repositories created in the last day alone include two different low-precision quantisations, a four step lightning variant, and hybrid builds mixing the model’s image-to-video and reference-to-video modes. On the forums somebody reports a LoRA getting 720p in minutes, somebody else is generating a clip a minute at low resolution, and there is now a node specifically for video outpainting with it.

Alongside that, the creative posts have started. A LEGO style short. An immersive environment piece. Somebody working out how to hold a consistent point of view across shots, which is the hard part of using any of these models for actual storytelling.

That progression, from can it run, to can it run fast, to look what I made, took about seventy-two hours.

The part that is ours

We have not run H3. We are going to bench it, for an episode rather than for production, and we said why yesterday: the licence puts an application form in front of anyone in the US, EU, UK or South Korea, which makes it a research subject for us rather than a tool.

But here is the number from our own work that we think matters more than the four hundred.

Earlier this month we benchmarked LTX-2.5, the update to the video model our studio actually runs, across two full rounds. Our conclusion was that we are not migrating. Our production still runs the previous version, because moving means rebuilding every graph we depend on, and our current setup is tuned and working. The ecosystem moved in days. We chose to move in months, on purpose.

Both of those things are correct at the same time, and the gap between them is the thing nobody tells you when a model trends. Four hundred repositories in five weeks is a real signal about how healthy open weights are. It is not a signal that you personally should rebuild your pipeline this week. A model that is genuinely better and a migration that is genuinely worth it are two separate questions, and the second one is answered by your own setup rather than by a download counter.

Also in the haul

NVIDIA’s DLSS 5 visual enhancer appeared as standalone neural rendering for images and video, and somebody has already wrapped it in an experimental ComfyUI node. Trellis.2 and Pixal3D both went native in ComfyUI, which is the 3D side of the house getting the same treatment. On the language model side, Qwen’s Flash Next line picked up multi token prediction support in GGUF form.

If you are trying to work out which of this your own machine could actually hold, the calculator is the quickest answer, and our licence routine is the two minute check worth running before you download any of it.

A daily note from our news radar. The week gets the full treatment in the DIY AI Brief every Monday.