Spaces:
Running
Running
metadata
title: README
emoji: ☕
colorFrom: purple
colorTo: pink
sdk: static
pinned: false
Humanjava Enterprises
Frontier AI you can hold, inspect, and run yourself.
We take models that only work inside someone else's walls and make them run in the open, on hardware you own. Every release ships with the recipe, so you can rebuild it, or build your own.
Releases
- DeepSeek-V4.1-Flash-1017-GSQ-RCO-upstream-GGUF: ISTA-DASLab's 2-bit quant of the 748B model, repacked byte for byte so it runs on stock llama.cpp. One 96 GB card, about 66 to 72 tokens a second.
- DeepSeek-V4.1-Flash-1017-obliterated-upstream-GGUF: the first model built end to end with our pipeline, an abliterated variant with fewer reflexive refusals.
Build your own
The repack tool and the build pipeline are open: deepseek-v41-upstream on GitHub. Bring any DeepSeek-V4.1-Flash checkpoint, a fine-tune, a merge, your own variant, and get a one-card file in the same size class.
Why we're here
The open engine everyone depends on is only as good as what people put back into it. This is our part. More in Why We're on Hugging Face.