I'm not quite sure about the parameters as ClaudeCode handles them automatically.
For now, I've turned off the thinking (reasoning) mode.
I downloaded both q8 and q6 but only tried running q8.
Without MTP, it gives about 50 TPS.
With MTP on, it goes up to around 150-160. It's amazing.
I compared Gemma4 with my local Qwen3.6-27b in a one-on-one match.
Qwen also gives about 55 TPS when MTP is on.
My power unit (PSU) is only 1000W, so I can't run both at full capacity and use them with a limit of 280W.
Gemma4 is definitely faster. The work speed seems to be more than three times faster.
However, the quality doesn't match Qwen's. When given the same prompt, Qwen often uses over twice as many tokens. If Gemma is like a student who finishes first and takes a nap during exams, then Qwen3.6-27b is like a student who works until the very end of the exam.
Coincidentally, a head-to-head video has been uploaded.
