🟢 All systems operational · MIT-licensed weights on hardware you rent
+ Add Provider
Open Weights · MIT licence

DeepSeek — MIT Weights,
Your Own Hardware

DeepSeek publishes its models under the MIT licence — code and weights both. R1 is a mixture-of-experts model with 671 billion parameters that activates 37 billion for any single token, and six distilled versions bring the same style of reasoning down to sizes that fit one card. Running it yourself keeps every prompt on your machine.

See what a GPU costs → Browse all GPU plans
deepseek — serve
$ ollama run deepseek-r1:32b
✓ reasoning model on a single card
$ vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
✓ OpenAI-compatible API on :8000
# the full 671B is a multi-GPU node,
# not a single rented card

What you get with the weights

MIT, including the weights
Code and model weights are released under the MIT licence, which permits commercial use and self-hosting. The distilled versions are built on Qwen and Llama bases and carry extra notes from those licences — read the one that matches the size you pick.
Mixture of experts
671 billion parameters in total, 37 billion active per token. The compute bill behaves like a much smaller model; the memory bill does not — all of it still has to be resident.
128K context
Long enough for a codebase or a stack of documents in one prompt — and the main reason real memory use runs above the weight figures below.
Six distilled sizes
1.5B, 7B, 8B, 14B, 32B and 70B, distilled onto Qwen and Llama bases. The 14B and 32B are the ones that turn a single rented card into a usable reasoning model.
The usual serving stack
Nothing bespoke: the weights load in the same engines everything else does, and the result speaks the OpenAI-compatible HTTP API your code already talks to.
Prompts stay where you put them
The reason most people self-host a model at all: what your users type never crosses a network you do not control, and there is no retention policy to read.

How much GPU memory each size needs

This arithmetic is ours, not a vendor requirement. DeepSeek publishes no per-GPU minimum, so we do the sum in the open: weights take roughly 0.5 byte per parameter at 4-bit, plus +25% for the KV cache and the serving runtime. It is a floor. A 128K prompt, a big batch or a different engine all push it up, and a mixture-of-experts model needs the whole parameter count in memory even though only a fraction of it does work on any given token.
Model Parameters Weights at 4-bit With headroom Fits
R1-Distill-Qwen 1.5B 1.5B ≈ 0.8 GB ≈ 1 GB Anything with a GPU
R1-Distill-Qwen 7B 7B ≈ 3.5 GB ≈ 4 GB An entry card
R1-Distill-Llama 8B 8B ≈ 4 GB ≈ 5 GB An entry card
R1-Distill-Qwen 14B 14B ≈ 7 GB ≈ 9 GB One 12–16 GB card
R1-Distill-Qwen 32B 32B ≈ 16 GB ≈ 20 GB One 24 GB card
R1-Distill-Llama 70B 70B ≈ 35 GB ≈ 44 GB Two 24 GB cards, or one 48 GB
DeepSeek-R1 (full) 671B ≈ 336 GB ≈ 420 GB A multi-GPU node

The “fits” column is a rule of thumb about memory only. It says nothing about how fast the answer comes back — that is the card’s bandwidth and generation, which our catalogue records unevenly, so we do not pretend to rank it here.

Cheapest GPU at each tier

All GPU plans →

The cheapest live GPU plan in the DohoHub catalogue at each memory tier, out of 142 plans whose provider publishes a GPU memory figure. Prices are the ones our feed tracks, normalised to be comparable and shown in the currency picked in the header. Where one machine is the cheapest way to clear several tiers, it is listed once and the note says which tiers it covers.

Provider Plan Specs Price Visit
Servarica
Bee GPU VPS
cheapest at 8–24 GB+ · 24 GB on board
 6 Cores24 GB vRAM30 GB RAM400 GB NVME100 TB traffic10 Gbps
$79.00/mo
 Visit
NovoServe
Supermicro X11 10SFF (GPU)
cheapest at 48 GB+ · 64 GB on board
 64 GB vRAM25 TB traffic
$355.99/mo
renews at $402.52
 Visit
GPU Mart
Enterprise GPU VPS - RTX Pro 6000
cheapest at 96 GB+ · 96 GB on board
 32 Cores96 GB vRAM84 GB RAM400 GB SSDUnlimited traffic1 Gbps
$649.00/mo
 Visit
NovoServe
Supermicro H12 4SFF+U.3 (8 GPU)
cheapest at 192 GB+ · 384 GB on board
 384 GB vRAM25 TB traffic50 Gbps
$3,138.73/mo
 Visit

We record the memory a plan advertises, not which card supplies it, and 48 GB spread over two older cards behaves differently from 48 GB on one modern card. The full 671B model wants roughly 420 GB even at 4-bit — more than any single listing here holds, so that one is a multi-node question these plans do not answer. None of these providers sells “DeepSeek hosting”; they sell GPU servers, and the install is yours.

Pick a card that fits the size you need

Filter the catalogue by GPU memory, cores, storage and location — live prices, tracked every six hours.

Find a GPU server →
↑