

In general, you take the model size in billions of parameters (eg: 397B), divide it by 2 and add a bit for overhead, and that’s how much RAM/VRAM it takes to run it at a “normal” quantization level. For Qwen3.5-397B, that’s about 220 GB. Ideally that would be all VRAM for speed, but you can offload some or all of that to normal RAM on the CPU, you’ll just take a speed hit.
So for something like Qwen3.5-397B, it takes a pretty serious system, especially if you’re trying to do it all in VRAM.
It’s not an either-or scenario. Running services in Docker/Podman is great and makes a lot of sense, as you’ve found. But there’s no reason the OS running those Docker containers can’t be a VM on a hypervisor like Proxmox. Then you get the simplicity of Docker, in addition to the isolation and segmentation (network and process) provided by VMs, and snapshot-based incremental backups from PBS. It’s the best of both worlds. You wouldn’t have a VM per service like you ran before, instead you’d have a VM per group of related services with common networking and security requirements. For example, all of your publicly exposed services can run in Docker in their own isolated VM that’s walled off from the rest of your network, while your internal-only services also run in Docker, but on a separate VM on your internal network.