Q4_K_M quality at 57% of the size
Helios compresses a trained model to a little over half the size of llama.cpp's widely-used Q4_K_M quant, at matching quality — a model that needed 4.68 GiB fits in 2.69.
See the benchmarksThe future is here.
At Nothing Inc., we are redefining the boundaries of intelligence, revolutionizing AI to create solutions the world has never seen.
Revolutionizing AI
The Helios format, by the numbers
Better than the aggressive quants
Against Q2_K, Helios is better on every accuracy instrument while landing a smaller file — and the result holds on a second model, not just the one it was tuned on.
Lossless, and local
Decompression is bit-for-bit exact on every run, and a single file runs at several memory and latency operating points — on the GPU you already own.
Applied
Curious how the research is being put to work?
Helios answers how much intelligence fits on a machine you already own. Float Point is where that answer becomes software — our research realized in real products.
See how it is applied“The compute required for AI training is growing exponentially — this is not sustainable unless we innovate on efficiency. We must figure out how to achieve more with less.”
— Sam Altman
We are currently looking to license Helios.
Stay tuned on our socials for more details — or reach out directly about business and licensing inquiries.
Get in touch