State of the art AI models like Claude, Grok or Codex are very powerful, but they get expensive fast. If you use them seriously, you’re looking at $100–300 a month, and sometimes more. The cheaper plans, which are all $20, burn through their limits quickly, and the higher tiers start to feel like too much spending when you’re using them every day for coding, writing, or agent work. Just by doing normal stuff you’re between $50-$100.
Open weights models have improved a lot in the last year. In many cases you can now get similar quality to the closed models, but at a fraction of the price — or completely free if you run them on your own hardware.
The catch is that choosing the right one is still confusing – even for me, if I’m being honest. There is no single company telling you “this one is good for coding, this one is good for agents, this one is too heavy for your machine.” You have to figure it out yourself.
I got tired of that, so I decided to make a short 4-episode mini-series that walks through the entire decision process. No theory for the sake of theory, just the basic things that actually help you pick a model and use it.
Episode 1 is already live and you can watch it from here. In it I cover:
- Why open weights models are finally worth serious attention right now
- The real difference between open source and open weights (most people mix these two up)
- The only four metrics that matter when you evaluate a model
- Parameter count and what the different size ranges are actually good for (1–8B, 14–70B, 100B+)
- Dense architecture versus Mixture of Experts (MoE), and why this changes local use requirements
- Quantization levels (2-bit, 4-bit, 6-bit, 8-bit) and the quality versus size trade-offs
- Text-only models versus multimodal models, and when each one makes sense The episode is short and focused. I kept it under ten minutes so you can watch it once and come away with a clear mental checklist.
The rest of the series will be in the same ballpark and they will cover:
- Episode 2: accessing models through third-party providers and APIs
- Episode 3: running models locally on your own hardware (and what the these hardware requirements look like)
- Episode 4: harnesses — the software layer that sits on top of the model and makes it usable, from simple command-line tools up to full agent interfaces
If you’re already paying for closed models and wondering whether open weights are good enough for your workflow, start with Episode 1. It gives you the basic language and the four criteria you need before you start downloading anything.
I’ll post the next ones as they go live.
Top comments (0)