Why We're on Hugging Face

If you follow our work, you have noticed we keep showing up in the same few places. The App Store. GitHub. Nostr. And now Hugging Face. It is not scattershot. It is the same thesis wearing different clothes: the tools that shape your digital life should be ones you can hold, inspect, and run yourself.

Hugging Face is where that thesis meets the models. So let me tell you what we put there, and why.

The thesis, in one line

The future of AI that most people are being sold is a meter running on someone else's machine. You rent the intelligence, by the request, and the terms are theirs to change. For a lot of work that is fine, and I have said so before. But some work should run on hardware you own, where the model cannot be altered or taken away, and where nothing leaves the building. That is not nostalgia for self-hosting. It is where the serious, private, high-trust work has to live.

The problem has always been that the best models were too big, or locked to an engine you do not control. That is the gap we have been closing.

What we actually shipped

ISTA-DASLab, a research group that does genuinely excellent quantization work, released a 2-bit version of DeepSeek-V4.1-Flash, a 748-billion-parameter model, squeezed down to where it can run on a single workstation. Remarkable work. But it shipped in the layout of their own proprietary engine, and the open engine almost everyone actually runs, llama.cpp, refused to load it. A frontier-class model, free to download, that most people could not run.

So we fixed the layout, not the model. Our repack tool rewrites the file into the structure stock llama.cpp expects. It does not requantize anything. Every quantized weight is copied across byte for byte, so what you run is ISTA's work exactly, now on the open engine, no fork and no special runtime. We put the ready-to-run weights on Hugging Face so you can download and go.

What that buys you, in plain numbers from our own machine: the model runs on one 96 GB card with about 138 GB in memory, at roughly 66 to 72 tokens a second, on nothing but upstream llama.cpp. A 748-billion-parameter model, answering at reading speed, on hardware a serious shop can actually own. That is the whole point made concrete.

Make your own, too

A repack is only half of it. We also published the build pipeline that takes any DeepSeek-V4.1-Flash checkpoint, a fine-tune, a merge, your own variant, all the way to a one-card file in the same size class. The first model we built that way is an abliterated version, where the reflexive refusals are reduced. To be honest about what that is: refusals reduced, not removed, capability the same as the base, and we have been plain in the repo about what we have and have not verified. It is there as proof the pipeline works on your own model, not as a headline. If you are going to own the model, you should be able to shape it, and then be accountable for what you shipped.

Why give it back

We could have kept the repack tool to ourselves. It is a real edge. We published it because the open engine that all of this depends on, llama.cpp, is only ever as good as what people are willing to put back into it. We have taken a great deal from that ecosystem. Contributing a tool that lets a whole class of locked models run on it is the rent.

That is why Humanjava is on Hugging Face, and on GitHub, and on Nostr. Not as marketing. As the same move every time: take something that only worked inside someone else's walls, and make it work in the open, on hardware you own. If that is the kind of AI your business actually needs, private, owned, and yours to run, the door is open.

Everything here is live: the tool and recipe on GitHub, and the weights on Hugging Face.