Most of the open vs closed talk here is about whether you can run the thing on your own hardware. fair, that's the obvious draw. but the part i think gets slept on is that open weights mean you can actually post train on top of the base, not just run inference.
With a closed api you're renting intelligence. you can prompt it, you can rag around it, but you can never make it yours. you cant fine tune the actual weights for your domain, you cant distill it down, you cant freeze a version and own it forever. You're permanently downstream of whatever the provider decides.
I saw some post about people post training their own models on top of glm-5.2 now that its open weight, and that framing stuck with me more than the benchmark numbers did. a frontier-ish base you can legally build on changes what a small team can do. You dont need to train from scratch, you start from something already strong and specialize it.
Realistically most of us arent fine tuning a 700b model in our basement, the compute is brutal and i wont pretend otherwise. but the option existing at all is the point. even renting cloud compute to post train your own variant is a completely different thing than being locked out of the weights entirely.
Anyone here actually post training on top of the bigger open models, or is it still mostly inference and the fine tuning stays in the small model range?
[link] [comments]