Member-only story
I still remember the high. ollama run llama3
, two minutes, and suddenly I had a model chatting with me on my own laptop. No CUDA nightmares. No three-hour Saturday lost to a build script. It just worked, and I was that annoying person telling everyone about it.
Then I tried to actually build something with it. That’s where the honeymoon ended.
It hides the good stuff. Ollama runs on llama.cpp underneath, which is genuinely powerful — but you barely get to touch that power. Sampling controls, quantization choices, batching, all the interesting knobs? Locked away or half-documented. It’s the “smart TV that hides the settings menu behind six sub-menus” problem, except the setting you want isn’t even in there.
**Ask it to handle more than one person, and it panics. **This is the real dealbreaker. Ollama was built for you, alone, one chat window. The moment a few teammates hit the same instance, things get ugly — queuing, latency spikes, memory creeping up. It’s not a server. It’s a very polite single-player experience wearing a server costume.
You never really know what you’re running. Quick story: when the DeepSeek-R1 distills dropped, a ton of people pulled the “R1” model through Ollama and thought they had the real thing. They didn’t — it was a Qwen or Llama model…