[d4rkp4ttern]

Curious what llama-server flags you used. On my M1 Max 64GB MacBook I tried it in Claude Code (which has a 25K system message) and I get 3 tps.

But with Qwen3-30B-A3B I get 20 tps in CC.