This, and depending on the workflow and usecase, you don't necessarily need the latest and greatest with the right kind of harness engineering.
Like everything in engineering, it's about tradeoffs and what works best for your specific problem.
Like everything in engineering, it's about tradeoffs and what works best for your specific problem.
Only since this year are most models robust in this way. In many open models, the problem is that successive generations are not that robust yet. i.e if you have a working setup, next generation of the same model family will need way too much reworking. So it's difficult to upgrade the model. Gemini is stellar in this respect (probably because, and I suspect, their flash models are distillations of the same larger model). OpenAI/Anthropic are too, but higher cost of newly released models is quite the blocker. Deepseek is especially hard to upgrade. Kimi doesnt target this segment, so people that can post train it do so. Qwen and llama are the stellar open ones, incredibly stable, although llama has stopped receiving upgrades in a while. It is still used a lot though.