Everyone keeps saying "open source AI" like it's obvious. For a while I nodded along and had no real idea what people were downloading.
So here's the plain version. There are two ways to use a powerful AI model, and the whole difference comes down to one question: who holds the model, them or you.
1. Closed models: you rent the intelligence
Claude, GPT, and Gemini are closed models. The model itself lives on the company's servers. You send a prompt, it sends back an answer, and you never touch the thing doing the work.
That is renting. Someone else owns the brain, pays the enormous compute bill to keep it running, and can change it or retire it whenever they decide to.
For almost everyone, this is the right call. It is the easiest, safest, and most capable option, and you never once think about hardware. I run my whole life on closed models and I am not switching.
2. Open-weight models: you download the actual model
A model's "weights" are the billions of numbers that make up its trained brain. That is the model, as a file. Open weight means the company publishes that file for anyone to download. Llama, Mistral, Qwen, DeepSeek, and now Kimi all do this.
Once those weights are on your hardware, they are yours. The model runs offline, no one can revoke it, and it never quietly changes underneath you unless you change it.
3. The catch nobody mentions: you pay the hosting
When you rent a closed model, the provider eats the compute bill. Download an open model and that bill becomes yours.
Running a model takes real hardware. The chips, the electricity, the setup, and the work of keeping it up. That is the actual trade for owning it. Free to download is not the same as free to run.
This is the whole reason closed models still win for most people. The download is the easy part. The hosting is the part with a price tag.
4. Three ways to actually host one
I think about hosting as a spectrum, from "most control, most work" on one end to "least work, least control" on the other. Here is the whole map:
Read left to right, here is what each one really means:
- On your own machine. Tools like Ollama or LM Studio run smaller open models, roughly 7 billion to 70 billion parameters, right on a good laptop or desktop. Private, offline, free after the hardware. This is where to start.
- Rent a GPU by the hour. Services like RunPod, Modal, or Baseten hand you a big machine in the cloud. You run any size model on it, and you pay only while it is switched on. More control, more setup.
- Use a hosted version of the open model. OpenRouter, Together, and Fireworks already run these models for you. You call them like any normal API, with no hardware to manage, and you can still switch providers any time. Easiest by a mile.
5. The one to watch right now: Kimi K3
The reason this whole topic went mainstream is a model called Kimi K3. Moonshot AI, a lab out of China, released it in July 2026, and it is the largest open-weight model anyone has shipped.
The numbers are genuinely wild: 2.8 trillion parameters, with 104 billion of them active on any given answer, a memory that holds a million tokens at once, and it reads images and video, not just text. It hit number one on Hugging Face within thirty minutes of release. The weights are free to download under the Kimi K3 License.
Because it is multimodal, you can hand it a picture and a mess of context and let it sort it out:
How to get Kimi K3
One honest thing first, because I would want someone to tell me. K3 is data-center scale. Self-hosting the full model needs something like eight H100 GPUs at the very minimum. So for almost everyone, "download it" is aspirational. Try the hosted version first, and only pull the weights if you have the hardware and a real reason.
Here is where each path lives:
- Try it free, no setup: the chat at kimi.com, or through a hosted provider like OpenRouter, Together, or Fireworks.
- Build on the API: platform.kimi.ai, using the model name kimi-k3.
- Download the actual weights: huggingface.co/moonshotai/Kimi-K3, pulled with the Hugging Face CLI or the transformers library.
- Read how it was built: the technical report at github.com/MoonshotAI/Kimi-K3.
What to actually do with all this
- If you just want the best AI, rent it. A closed model like Claude is the right answer for almost everything you will ever do.
- Open weight is about control, not saving money. You reach for it when you need privacy, offline access, or to own the thing outright.
- Start hosted, not local. Try an open model through a provider before you even think about buying or renting a GPU.
- Downloadable is not the same as runnable. The bigger the model, the bigger the bill to keep it on.
Closed or open is not better or worse. It is rent versus own, and now you know which question you are actually answering.
If you are still deciding between the big closed models before you go anywhere near an open one, here is my honest six-month comparison of ChatGPT and Claude.