Why doen't AI companies agree on a single standard?
retained text
I saw that models perform wildly different based on what prompts they use, their inference parameters, the harness, the type of quantisation used, context length, the kind of inputs they take etc. So, why don't companies agree on a set of standards? Like "this is the harness we'll use", "this is the quantisation type we'll use", "this is the prompt style we'll post-train on" etc. It'd be beneficial having a set level of standards, and enough freedom for the labs to improve their own models. Its not like these will affect them. I'm new to local models after using cloud models for years which "just works". I hate how many different engines, api formats, harnesses, tools I had to go through to optimise my setup. Maybe I'm missing something.
document fetched 2026-09-25 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1wp6loz/why_doent_ai_companies_agree_on_a_single_standard