The Mac Studio Ultra was too expensive, so I only looked at it... but I thought this won't do, so I hastily purchased a DGX SPARK.
Issues:
I was using Codex Desktop installed on MacBook Pro (24GB) for automation and quant market analysis, and it worked well, but since this is something I need to carry around, I kept it plugged in while doing scheduling work, which feels like such a waste, right? I was using the MacBook Pro like a Mac Mini.
Installation:
What struck me was how much better things have become - I just opened one terminal, clicked on Codex Desktop, and it set everything up for me. I just open a terminal for Codex, give it the NVIDIA homepage, and say "install Hermes, vLLM, and WebUI" - and it's done in one go.
So in just one day, I installed and ran Qwen3.6 and Hermes automation, and the token speed was about 100T/s or more... slightly impressed, and then impressed once more after connecting Qwen to Claude Code. And then I had Hermes bring over all the MCP I was using with Codex and transplanted it all, and that was it.
Impressions:
I haven't found an AI model that can surpass Qwen3.6 yet. Everyone must be disappointed!
Hermes is so easy to install. Actually, it's not an easy task, but it feels like a doctor healing itself.
It seems like locally installed AI is stagnant at a mediocre level, but Quant doesn't need high-level reasoning like that, so I actually prefer a stable environment like this.
By the way, Codex was difficult because it didn't listen to instructions properly (it calculated differently from what I asked, and even when I pointed it out, it insisted it was correct), but Qwen coolly admits when caught cheating and moves on, so productivity is actually better. -_-
Conclusion:
If you want deep AI reasoning, just go with the cloud. But for repetitive bulk calculations and task work, etc., if cloud costs are insufficient even with a $200/month plan, it's definitely worth considering. With 128GB, the limit is up to 120B, so if you really want high-level reasoning, it's not the right choice.
