everything will be local eventually
retained text
I know current llms like opus 5.5 are in a league of their own. but we all know they're losing a shti ton of money with their subscriptions. ALSO we all know how tasks like OCR, segmentation etc are getting solved by single digit B parameter models. I'm not saying llms are a gimmick but I'm saying the future is already shaping up to be about specialized models that work better when they're smaller. easier training. so datacenter level power not needed. ten years from now we'll have a specialized model for the tasks we care about and llms will fail to do everything all at once (unless they keep calling it one model but have a hundred under the hood). system one models are a bad but workable example. you could tell an llm to rate options all day but get yourself on 9B parameter model trained by some lab and it tops the benchmarks.
document fetched 2026-10-24 10:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1xdcpxb/everything_will_be_local_eventually