🟢 All systems operational · Apache 2.0 weights on hardware you rent
+ Add Provider
Open weights · Apache 2.0

GPT-OSS — OpenAI’s
Weights on Your Rack

OpenAI publishes two open-weight models under the Apache 2.0 licence: gpt-oss-120b and gpt-oss-20b. Both are mixture-of-experts, both ship already quantized, and — unusually for this section — both come with a memory figure stated by the people who trained them. That makes picking hardware a lookup rather than an estimate.

See what a GPU costs → Browse all GPU plans
gpt-oss — serve
$ ollama run gpt-oss:20b
✓ 14 GB pull, 128K context
$ vllm serve openai/gpt-oss-120b
✓ OpenAI-compatible API on :8000
# 120b is stated to fit one 80 GB card,
# 20b within 16 GB — by OpenAI, not by us

What OpenAI actually released

Apache 2.0, weights included
A plain permissive licence with no revenue threshold, no field-of-use clause and no acceptable-use appendix bolted on. Among open-weight releases that is the easy end of the licensing spectrum, and it is worth noticing.
Two sizes, both sparse
gpt-oss-120b is 117 billion parameters with 5.1 billion active per token. gpt-oss-20b is 21 billion with 3.6 billion active. The compute behaves like a small model; the memory bill does not, because every parameter still has to be resident.
Quantized at the source
The mixture-of-experts weights are post-trained in MXFP4, so the download you get is already the compact one. This is precisely why a 117-billion-parameter model has any business being on a single card.
Three levels of effort
Reasoning effort is configurable at low, medium or high — described by OpenAI as fast responses for general dialogue, a balance of speed and detail, and deep analysis respectively. The same weights, cheaper or dearer per answer.
Runs in what you already use
OpenAI names Transformers, vLLM, PyTorch with Triton, Ollama and LM Studio as ways to run it. There is nothing bespoke to install, which is why this page can point at the same GPU plans as the rest of the section.
Nothing goes back to OpenAI
These are weights on your disk, not an API key. Once the download finishes there is no upstream call, no rate limit and no terms-of-service change to keep track of.

The memory figure, published by OpenAI

For once we do not have to estimate. Every other model page in this section does its own arithmetic, because the vendors publish parameter counts and leave the hardware to you. OpenAI is the exception: it states that gpt-oss-120b fits into a single 80GB GPU, naming the NVIDIA H100 and AMD MI300X, and that gpt-oss-20b runs within 16GB of memory. Those are their numbers, not ours, and they are only possible because the weights are already MXFP4. Serve the same models at a higher precision and you are back to doing the sum yourself.
Model Parameters Active per token Memory stated by OpenAI Download size
gpt-oss-20b 21B 3.6B within 16 GB 14 GB
gpt-oss-120b 117B 5.1B a single 80 GB GPU 65 GB

Both models carry a 128K context. The download sizes are the ones Ollama lists in its own library, and a download size is not a memory requirement — the context and the runtime sit on top of it. On performance we report only what OpenAI claims about its own models: that gpt-oss-120b reaches near-parity with o4-mini on core reasoning benchmarks and that gpt-oss-20b performs similarly to o3-mini on common ones. DohoHub does not run benchmarks, so those figures are theirs and are labelled as such.

Cheapest GPU that clears the bar

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. The two tiers that matter here are the ones OpenAI names: 16 GB for the 20b and 80 GB for the 120b. Prices come from our own feed, normalised so they are comparable and shown in the currency picked in the header.

Provider Plan Specs Price Visit
Servarica
Bee GPU VPS
cheapest at 16–24 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 80–96 GB+ · 96 GB on board
 32 Cores96 GB vRAM84 GB RAM400 GB SSDUnlimited traffic1 Gbps
$649.00/mo
 Visit

“Within 16 GB” means the weights and a working context, not the weights and a crowd — if several people will use it at once, take the next tier up and read the vLLM page on why. We record the memory a plan advertises, not which card supplies it, and 80 GB spread over two cards is not the single 80 GB GPU OpenAI describes. None of these providers sells “GPT-OSS hosting”; they sell GPU servers, and the weights are a download away.

Pick a card that clears the published figure

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

Find a GPU server →
↑