Turned on my Spark again.
This machine. Got Qwen3.6-35b running. I really love this model. The intelligence density per gigabyte on this thing is bonkers, and I'm more convinced than ever that this is only a sign of things to come on the local agent front.
The code it generates is well organized, extremely well written for a local model, and it can handle very complicated coding tasks and design processes without losing the thread.
I totally understand why people are excited about it.
Ran it under ollama with the base configuration, and the thing just screamed. Then I hooked it up to Ollama Codex, and no problems... for about an hour.
The hardware experience
But the DGX Spark makes me nervous.
Let's recap my experience with the Spark so far.
Ordered one three days after it was released. Was thrilled, excited, and happy to play with it. It was one of those value propositions, where, assuming it lived up to its promises would have cut about $11,000 in annual AI subscription expenses for me.
So the spend seemed... totally justifiable?
Then I got it. And when I got it, I got busy developing with it.
New problem: It runs hot. Very hot.
I didn't think about this, because the expectation is that you leave computers on for weeks or months at a time without stopping them.
Started it on a Monday, let it run building things until Friday. That Friday... it got so hot that it melted a hole in the vinyl cover of my shiny new standing desk, which I had purchased only three weeks before. It had this smell to it.
And it would start after that, but it wouldn't stay on for more than a minute or two without rebooting.
Wrestled with Nvidia support for two months, and finally got a replacement.
I was reluctant to turn it on.
The new one takes longer to get searing hot, and doesn't seem to be getting hot enough to actually burn my skin. It's more cup of hot coffee hot, than space heater hot like the last one was.
And the failure mode is a lot more pronounced.
The last one ran hotter, but the system stayed coherent for longer.
On this one, the models just stop responding after about an hour. But... I'm less worried about burning down my house with it.
So... progress?
Have any of you had any luck on cooling this thing down?