Hi Andrey, I bought a new PC when memory was still cheaper, and without graphics card, but AMD 8700G CPU and 64 G RAM, I can use several models purely local. gpt-oss gives me 15-20 tokens per second, converting sources. GLM is Ok too but slower, but I found the Qwen models too chatty. But there might be tuning setting that improve that. I use a terminal interface using AMD's lemonade, not yet a local agent like pi. MatN On Tue, 22 Sep 2026 09:50:59 +0300 Андрей Спицын via Cin <cin@lists.cinelerra-gg.org> wrote:
Hi Andrew
My other problem with AI (apart from resource use and how it shoveled into
everything) is who control it
This is indeed a problem if you use the Claude code or Codex. However, if you have a high-end gaming graphics card, such as an Nvidia 3090, you can run local AI models such as the Qwen-3.8-27B model. I have had very good results with it.
Best regards, Andrey