Open-Weight vs Closed AI Models: Which Should You Actually Use?

Every few weeks someone posts the same fight online: open models vs closed models, as if one side is clearly winning. I used to pick a side too. Then I started shipping real things with both, and the fight stopped mattering. What matters is the job in front of you.

This is not a philosophical essay. It is how I choose, after a year of building AI projects from Kathmandu and breaking things in public. I am not loyal to a license. I am loyal to whatever gets the job done and keeps me owning the result.

First, the terms people mix up

"Open source" and "open weight" are not the same thing. Open weight means the trained parameters are published, so you can download and run the model yourself. The training code and the data often stay private. That is Llama, Qwen, and DeepSeek. Closed means the model lives behind an API you call, like ChatGPT, Claude, and Gemini. You never touch the weights.

Both can be "free" to start. The difference is who controls the model after you depend on it.

When I reach for a closed model

If the task needs the absolute best reasoning, the longest memory, or the smoothest multimodal input, I use a closed frontier model through its API. For client work where reliability beats everything, I am not going to gamble on a smaller local model.

Closed wins on three things:

• Quality ceiling. For hard reasoning, frontier APIs are still ahead.
• Zero ops. No GPU, no server, no updates. You call it and forget it.
• Speed of shipping. When the deadline is tonight, the API is the fastest path.

The cost is real though. You pay per token, your data leaves your machine, and the model can change or get more expensive without asking you.

When I reach for an open-weight model

This is the part the hype posts skip. Open weight is not "worse for cheap people." It is the right call when any of these are true:

• Privacy. Medical, legal, or user data that cannot leave your infrastructure.
• Cost at scale. If you serve 2 million requests a month, a one-time GPU bill beats a per-call API forever.
• Control. You want to fine-tune, quantize, or pin a version that never drifts.
• No internet. Offline tools, embedded devices, or a flaky connection.

Running something like Qwen or Llama locally means the model is yours. It will not get "upgraded" underneath you at the worst moment.

The real trade-off nobody sells you

Closed models feel free until the bill arrives. Open models feel free until you pay for the GPU, the engineer time, and the latency work. Pick based on which bill you would rather owe.

I do not ask "which model is best?" I ask "which model lets me ship this specific thing today, and keep owning it tomorrow?"

My default setup

For most of my projects: a closed frontier model for the hard first draft and reasoning, then an open-weight model self-hosted for the bulk, repetitive, private work. One does the thinking, the other does the volume.

You do not have to marry either. Download a weight, try it for a weekend. If it is not good enough, delete it. That option is the entire point of open weights.

The honest answer

Use closed when you need the best brain and fast shipping. Use open weight when you need privacy, scale economics, or control. Most serious builders end up using both, and the people shouting that one side won is just marketing with extra steps.

Start with the API. When the bill or the privacy question bites, that is your signal to go open weight.