HN Today there’s a small model called Kev. It comes in three sizes: 0.8B to 9B. Just stack it on top of Tongyi Qianwen 3.5, with the goal of letting you train and run decision-type models on a home GPU.
The model weights are公開 on Hugging Face, and the training scripts and evaluation datasets are packaged together.
This track is beneficial for encryption. Running small models locally, it’s the same path as small-compute nodes you’d want to build with proxy protocols like $FET . Training no longer gets stuck behind the big companies’ few-thousand-GPU bottleneck; individual users’ devices can run it, and proxy protocols can finally have real deployment scenarios. $NEAR is also related—this “subnet narrative” has been waiting all year for a local inference breakthrough.
The question is: are 900 million parameters really enough? It still depends on real evaluations. For open-source models, what I care about more is whether they can replace closed-source models on small tasks—not the parameter count itself.
#开源AI #代理协议 #Technology
The model weights are公開 on Hugging Face, and the training scripts and evaluation datasets are packaged together.
This track is beneficial for encryption. Running small models locally, it’s the same path as small-compute nodes you’d want to build with proxy protocols like $FET . Training no longer gets stuck behind the big companies’ few-thousand-GPU bottleneck; individual users’ devices can run it, and proxy protocols can finally have real deployment scenarios. $NEAR is also related—this “subnet narrative” has been waiting all year for a local inference breakthrough.
The question is: are 900 million parameters really enough? It still depends on real evaluations. For open-source models, what I care about more is whether they can replace closed-source models on small tasks—not the parameter count itself.
#开源AI #代理协议 #Technology