Considering new computer options
Weblog notes from June 27, 2026 that were compiled and shared
Earlier today I spent some time researching buying an Apple Mac Studio, but based on supply pressures or the impending M5 refresh later this year the shipping date seems to be October. I knew that demand for the Apple Mac mini had spiked with people wanting to install OpenClaw on them and run autonomous agents. The Hermes Agent from Nous Research seems to work well on my MacBook Air, but I want to switch over to Studio or mini to really utilize all the features. It turns out the Hermes Agent is just not as integrated with Windows and I strongly prefer using it on my MacBook. The only problem is that the MacBook Air cannot support any large Ollama driven workloads given its thermal regulation limitations. In hindsight, I probably should have gotten a MacBook Pro. That is something to consider for the next upgrade cycle.
My other option would be to upgrade the graphics card on my desktop, but that does not really solve the integration elements of the agent that I prefer within the Apple ecosystem. I know that the Hermes Agent development is primarily driven within that environment and that explains the integration being more useful. I decided to pay for some tokens from the Portal ecosystem Nous Research runs to get a feel for how the Claude Opus 4.8 model performs. To be fair it really is awesome. Going back to using the Qwen models locally using Ollama really does feel like a step backward. After you get a feel for how the cutting edge models perform it is easy to see that the next generation of models are going to be pretty epic.
It’s interesting to see how the model generations are translating to how much VRAM they need to run effectively locally using something like Ollama. A lot of the graphics cards that were sold for gaming cap out at 8 gigabytes of VRAM which means when the models shifted to being larger in terms of parameters a generation of graphics cards got left behind. The interesting part of that equation is that with the 8, 10, 12, 16, 24, and even 32 gigabyte cards they are going to quickly get outpaced by the next generation of models. A researcher that wants to use one of the models locally can probably be successful to get away with something above 16 gigabytes for now, but based on my usage of Claude Opus 4.8 the frontier models are going to quickly outpace what can be done with regular graphics cards. It’s an interesting situation where the next generation of models are not going to be able to run locally for most people.