So by now you've probably heard the noise. Some Chinese AI lab nobody outside the field had heard of a month ago put out a model called R1, and on Monday it apparently vaporized something like $600 billion off Nvidia's market cap in a single trading day. My brother-in-law, who owns exactly one stock and it's Nvidia, texted me three separate times asking if he should sell. I am not a financial advisor. I told him to ask literally anyone else.
But that's not really what I want to write about, because every tech site on earth already has five posts up about the Nvidia crash and the App Store rankings and whether this proves you can train a frontier model for $5.6 million (probably not exactly that, the number is doing a lot of work, but sure, roughly "way less than everyone assumed"). What I actually did this weekend, before any of that market stuff happened, was just... run the thing myself. On my own laptop. Because the DeepSeek app climbing to #1 ahead of ChatGPT made me curious, and also mildly uncomfortable about sending my prompts to servers in Hangzhou, so I figured I'd cut the app out entirely.
Getting it running is genuinely easy
I use Ollama for local models already (mostly for messing around, not anything serious), and they'd already added DeepSeek's distilled versions to their library within a day or two of release. The full R1 is a 671-billion-parameter beast that needs hardware none of us have sitting in a bedroom. But DeepSeek also released a handful of smaller "distilled" versions, trained on R1's outputs and built on top of existing Qwen and Llama architectures, ranging from 1.5B all the way up to 70B parameters.
On my 2021 MacBook Pro (M1 Pro, 16GB RAM, which is starting to feel a little embarrassing next to what people are running now) I pulled the 7B version first:
ollama run deepseek-r1:7b
That's it. No account, no API key, no app to install. It downloaded about 4.7GB and started answering questions maybe ninety seconds later.
What surprised me
The thing that actually got my attention wasn't the answers themselves, it was watching it think. R1 puts its reasoning in a separate block before the final answer, wrapped in what's basically a visible scratchpad, and reading it argue with itself is weirdly compelling. I gave it a logic puzzle about scheduling three people across overlapping meetings and it talked itself in a circle for a good four paragraphs before landing on the right answer. Kind of like watching someone do long division out loud.
I bumped up to the 14B version after that and my laptop fan, which I swear had been silent since roughly 2023, spun up like it was trying to achieve liftoff. Responses got noticeably slower too, maybe 15-20 seconds for a decent-length answer versus a few seconds on the 7B. Worth it for harder stuff, not worth it if I just want a quick answer while I'm making coffee.
Quality-wise: for coding questions and math it held its own against models I'd normally reach for, though I wouldn't call it dramatically better than what I already had access to. For anything touching Chinese politics or history it either goes vague or refuses outright, which nobody should find surprising given who made it, but it's worth knowing before you rely on it for anything remotely adjacent to that territory.
I'm not going to pretend I have some grand take on what this means for Nvidia or for OpenAI's pricing or for the next eighteen months of AI investment. People smarter than me are arguing about that right now on every podcast feed I subscribe to, and I'll admit I've started skipping past those episodes because it's the same four opinions rearranged. What I do know is that running a genuinely capable model, for free, entirely offline, on a laptop that's three years old, felt different than reading about it. It took maybe ten minutes to set up. If you've got Ollama installed already there's no excuse not to just try it yourself instead of taking anyone's word, mine included, for how good it is.