In Episode 1 I covered the four metrics that actually matter when you evaluate an open-weights model: parameter count, architecture (dense vs MoE), quantization, and whether the model is text-only or multimodal.
Once you know what to look for, the next practical question is: how do you actually use the model?
There are two main ways.
You can download the weights and run the model yourself, or you can use a third-party provider that already hosts it for you. This episode is about the second option — the providers.
When a new open-weights model is released, a bunch of companies that already have data centers and spare compute quickly make it available. You don’t have to download anything. They host the model and expose it through an API, usually in the OpenAI style. That means almost every tool, library, or agent framework you already use can talk to it without major changes.
The biggest practical advantage is convenience. With a single API key from one provider you usually get access to a whole set of models: DeepSeek, Kimi, MiniMax, GLM, Qwen, and others. You can switch between them depending on the task instead of managing separate accounts and keys.
Pricing is almost always pay-as-you-go and split into two parts: you pay for the input tokens (what you send to the model) and for the output tokens (what the model generates). This becomes important with models that “think” a lot or produce long responses, because the output side can add up quickly. Still, even with that, the cost is usually a fraction of what you pay for the big closed models.
Most providers also do some form of intelligent routing in the background. They automatically send your request to the instance that has the best availability or the lowest latency at that moment. You don’t have to manage that yourself.
This approach works especially well when:
- You have reliable internet
- You’re working from a thin laptop and move around a lot
- You don’t want to carry (or buy) a machine with 64 or 128 GB of RAM just to run models
- Your work spans different categories — coding one hour, writing the next, maybe some audio or video tasks later — and you want to switch models easily
- In short, third-party providers give you the convenience of closed models while still using open weights, usually at a much lower cost.
Of course this is only one side of the story. The other side is running the models yourself on your own hardware — either on a laptop or on a small server under the desk. That’s what Episode 3 will cover, including the realistic hardware requirements and the trade-offs involved.
And if you haven’t seen Episode 1 yet, start there: https://youtu.be/iBOh7atzUZY
I’ll post the next ones as they go live.
Top comments (0)