datapoint: 2x M5U cluster (GLM/Qwen)

Caveats: IDK how to set anything up and I asked Deepseek to handle it - not enough training data so without specific guides the set up doesn't work well. I'm not a programmer. My previous model was Qwen 3.8 27B q4 using ninfer. Was getting around 1.4k prefill and 100-120 tokens decode. Just want to provide data for anyone interested in purchasing M5U 256gb (80 cores) x 2. And if you have suggestions on how to optimize my set up I'd really appreciate it. GLM 5.3 q6 cluster: prefill 4.9k (idk if this is right but this is what the LLM says, two thunderbolt 5 cables), decode: 50 tok/s. oMLX set up with Deepseek API. Qwen Flash Next q6 cluster: prefill 8k and decode 200 using Tensorfold. I copied their guide, sent it to Deepseek, and asked it to set it up. GLM was too slow for me so I am using Qwen Flash Next. I am using Hermes as the harness (hermes has deepseek harness as research agent and Pi harness for coding agent). I am wondering whether using a different harness with less context overhead would make things faster or if since it is mostly context it doesn't really make a difference.