Upgraded my local setup with 2 rtx pros and it's amazing.

Follow up post of . Thanks everyone who replied. I didn't change the specs. Might be loosing some of the memory bandwidth but will scale in future if I need to. It took me 2.5 days to build it because one of the GPU connected to PSU was loosing power whenever I load anything on the GPU, so I had to rewire every connection again to identify the fault. I am glad the system is working because I was apprehensive if this will work (I am software dev, getting my hands dirty with hardware for the 3rd time in life). My finger tips still hurt from pulling the cables from motherboard and PSUs. To enable the full potential of the system, I had to enable peer to peer communication between the GPUs, cuda graph, tensor parallelism. I have capped both the GPUs at 500W (no reason, just didn't want GPUs to run on its full capacity). Also, I had to open my box, because temps were shooting high, and fans were making weird noises. I am running: Qwen 3.8 flash next 8 bit Deepseek v4 flash 0731 (official) I have a M3 ultra 512, LLMs run on it, but I personally find it useless for inference. My head just hurts watching it work slow. On the other hand this new system is killing it, decode 150 tk/s and prefill 10K tk/s. Qwen is good, but most of the context is consumed by thinking tokens, I was checking if it's a good idea to not use the thinking token. I barely have vram left for concurrent requests with full context window. Loving the Deepseek 1M context, and I also have room for 4 concurrent requests. Both of them are okay model, even if they make mistakes, I don't notice because of the speed. It's just fast, makes an error, corrects it moves on. Finally the day is here when I can save on monthly subscriptions and not worry about the weekly or 5 hours limit. I have already setup my server with openclaw, opencode, openweb UI and Tailscale. Has anyone experience excluding the thinking tokens of Qwen from the context and keep the final result? Was there any impact on the performance or accuracy of the model? Any suggestions, what else I should install on it? Any new models to try?