The cheapest realistic self-hosted setup for DeepSeek V4 Flash is not a single consumer GPU; it needs around 345 GB VRAM at FP16 , while an INT4 quantized setup drops to about 86 GB VRAM. A published GPU recommender says the cheapest fit for FP16 inference is 4× RTX PRO 6000 96GB at about $3.60/hr total , and INT4 can run on a much smaller setup, with RTX PRO 6000 96GB at about $0.90/hr listed as the cheapest per-GPU option there.

Cheapest way to host it

If your goal is “cheapest possible,” the practical order is:

  1. INT4 quantization on one high-VRAM GPU or one smaller multi-GPU node if your quality tolerance is okay.
  2. FP16 on 4× 96GB class GPUs if you want closer-to-original behavior.
  3. Cloud rental instead of buying hardware unless you’ll use it heavily enough to justify capex.

The guide I found also frames self-hosting as mainly an inference-only problem, meaning you are paying for weights plus KV cache, not training.

What that means for cost

At the cited pricing, the rough floor is:

  • INT4: about $0.90/hr for a 96GB-class GPU option.
  • FP16: about $3.60/hr for a 4-GPU 96GB setup.

That means monthly cost can be very different depending on usage:

  • 24/7 at $0.90/hr: about $648/month.
  • 24/7 at $3.60/hr: about $2,592/month.

If you only need it occasionally, renting wins. If you need constant high usage, buying hardware may eventually be cheaper than paying hourly, but the upfront cost is much higher.

Cheapest practical recommendation

For most people, the cheapest sane path is:

  • Use INT4 quantization.
  • Run it on a single 96GB GPU if possible.
  • If you need FP16 quality, use a 4× 96GB setup.
  • Keep the server lean: Linux, vLLM or similar inference stack, minimal extra services.

My take

If you want the absolute lowest spend, don’t aim for “perfect” self-hosting. Aim for a quantized model on rented hardware , then only move to owned GPUs if your usage is steady enough to amortize the purchase. For DeepSeek V4 Flash specifically, the numbers I found suggest the jump from cheap-ish INT4 to full FP16 is massive, so the quantized route is the only realistic low-cost starting point.