Comment by syntaxing

I use it with a strix halo server. 35B runs stupidly fast. 27B is about 700 TPS prefill and 30 TPS token generation. Which interestedly is about what Kimi K3 gives me depending on provider.