Woke up yesterday to the TV in the other room going on about some half-trillion-dollar AI infrastructure announcement, Trump at a podium with Sam Altman and Larry Ellison and Masayoshi Son lined up behind him like a boy band. Stargate, they're calling it. $500 billion for data centers over the next four years, supposedly. I had my coffee, thought "sure, fine, we'll see," and went back to what I'd actually been doing all weekend, which was trying to get a much smaller and much weirder AI thing running on my own laptop.
That thing is DeepSeek R1. A Chinese lab put it out Monday, dropped the weights right onto Hugging Face under an MIT license, no waitlist, no API key begging, just here you go. And unlike basically every other "reasoning" model since OpenAI's o1, you can actually download this one and run it on hardware you own. I've been messing with it since Monday night and I'm still not totally sure what to make of it, but I like it enough that I want to write down how I got it going before I forget my own steps.
What you actually need
I'm running this on an M1 Pro MacBook with 16 gigs of RAM, which is nothing special, and it matters a lot which size model you pick. DeepSeek didn't release just one model, they put out a family of distilled versions at 1.5B, 7B, 8B, 14B, 32B, and 70B parameters, plus the full 671B monster that basically nobody outside a data center is running. The trick is these smaller ones are distilled from the big reasoning model down onto Qwen and Llama base models, so you get a decent chunk of the reasoning behavior without needing eight H100s in your closet.
The easiest path in is Ollama. If you've already got it installed (and if you don't, ollama.com, five minute install, I'm not walking through that part):
ollama pull deepseek-r1:14b
ollama run deepseek-r1:14b
That pulled down something like 8-9 gigs for me and took about ten minutes on my connection. The 7B and 8B versions are faster to grab if you're impatient or your disk is already full of other models you downloaded in a fit of enthusiasm and never opened again (guilty).
The thinking-out-loud thing
Here's what actually got me hooked. When you ask R1 something, it doesn't just answer. It opens with a <think> tag and just narrates its own reasoning, sometimes for a paragraph or two, second-guessing itself, backtracking, going "wait, that's not right" mid-thought, before it closes the tag and gives you the actual answer underneath. I gave it a fairly annoying probability question I usually use to poke at these models and watched it talk itself in a circle for a bit before landing on the right answer. It's a little unsettling to watch, honestly, like reading someone's diary while they solve your homework.
The 14B model is my sweet spot on this laptop. The 32B is genuinely sharper, noticeably so on anything with actual multi-step logic, but it turns my laptop into a space heater and a fan louder than my dishwasher, and a single response can take a couple minutes to finish thinking before it even starts answering. Not worth it for day to day use unless I'm doing something I actually care about getting right. The 1.5B model, meanwhile, I'd skip entirely unless you're just testing that the pipeline works, it hallucinates its own reasoning half the time and I don't trust a word it says.
Is it actually as good as o1
For the stuff I tried, code explanation, some logic puzzles, one annoying regex I'd been avoiding writing by hand, the 14B and 32B distills held up better than I expected against o1-mini, and I paid nothing and sent none of it over the network to anyone. That second part matters more to me than people probably think it should. I'm not saying it beats full o1 on hard math, I haven't tested that rigorously and I don't fully trust my own gut on it yet. But for a Monday release with weights just sitting there for anyone to grab, it's a lot more than I expected to be writing about this week.
Anyway. Stargate can build its half-trillion-dollar data centers. I've got a MacBook that sounds like it's about to take off and a model quietly arguing with itself in a terminal window, and that's plenty for me tonight.