Selfhosted

Depends on how much quantization, but still fairly beefy, couldn't run it on my homelab with a 3080ti for example.

I generally use smaller 8-12b models and they're alright depending on the task.

I'm running 2x4090, the 35B fits very comfortable in that.

For large models like the 397B without a ton of money there are several ways, ive seen posts of people using arrays of used 3090s with good results.

The other option is CPU inference although with current RAM prices that is less cost effective.

I was looking at maybe an array of Milk-V JUPITER2 since vllm added riscv support which could be very cost effective.

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.