Buy a card. Run your own. It'll be some time to optimize it, but an AMD R9700 is one of the cheapest by $/vram. Sadly, its been increasing in price slowly, and

Buy a card. Run your own. It'll be some time to optimize it, but an AMD R9700 is one of the cheapest by $/vram. Sadly, its been increasing in price slowly, and there seems to be some inventory issues, but I've switched to it exclusively. I hope to get a 2nd card to run other workloads simultaneously. I've been running Qwen27B Q6 with 130k context at 30tps. It's not bad. I think with an "autopilot" mode in pi, and sub agents to handle context window management better, I'd envision you could get close to unattended workflows. I haven't quite gotten that far though. Negative is you have to do it all, and the temptation to tinker is real.