Seinari
Frontier AI models, small enough to serve on one GPU.
Ternary models store every weight as −1, 0 or +1, which makes them about 2.5 times smaller than the usual 4-bit releases. Seinari writes the inference engines that run them fast, on data-centre GPUs, gaming cards and phones, and is building a cheap way to convert large open models to ternary.
The aim is the cheapest serving of frontier open models.
Chat with a 27B modelPrismML's Ternary Bonsai 2 27B on Seinari's engine, one RTX 5090
Website builderDescribe a page; the 27B writes it while it renders live
On an iPhoneA ternary 2B chatting in real time, entirely on the phone
On your Android phoneInstall the app: the 2B runs on the phone, on its fastest engine