[frogperson]

I dont think open weight models are going to be a big disrupter. At even a small scale you need a machine with 8 gpus just to run the larger models that still arent that close to SOTA. A machine like that is half a million per year. You might need a second machine for your dev environment or do fine tuning. Then you will need to hire an AI engineer who know who to tune the model. You'll probably need k8s and Dynamo expert to run infra to do the inference.

It all gets very expensive, very quickly. Sure some customers will pay this, but many, many more are just going to pay per token. It's the same model that made the hyper-scalers like AWS into giant cash cows. Only this time, tokens can be sold to both enterprises and grandma. Grandma never bought AWS, which means the market for convenient tokens is huge.

[ST33LDI9ITAL]

It'll get to point to where good enough for average user models will be able to be ran on common hardware. Agentic platforms are already moving towards smaller ,more efficient, and faster targeted models per agent/task/tool. Takes a bit more to setup but works great.

All your general knowledge frontier models are just going to be baked into every OS and browser. Coding is almost to point of being viable for self host.. I think the only real use for high end paid models is gonna be kinda like things now... high end scientific workloads.. professional rendering.. especially 3d and video, financials, etc.

A $4,000 mac can take you pretty far right now.. and we have the upcoming dgx hitting mainstream consumers. Give it another couple years. It's changing fast.

[verdverm]

As we look out further, local Ai will become even better. I run a DGX Spark with qwen36moe and it is between gemini-2.5 & gemini-3 capabilities. It uses around the same electricity as an incandescent light bulb, no water, token freedom.

[ChrisLTD]

I’m not sure the parent comment is talking about running the open weight models on your own hardware