Shared AI models: one API, every machine
Open AI models served by graphics cards across the whole pool, behind one OpenAI-compatible endpoint. Send a chat call with your account key and get an answer — no GPU to choose, no instance to start. You pay per token, counted on our server.
Every model starts on a machine in the pool when you call it.
The models
| Model | Good for | Context | Input, per million tokens | Output, per million tokens | Now |
|---|---|---|---|---|---|
Bonsai 2 2Bbonsai-2:2b | The smallest: ternary weights that fit any card. Short answers, sorting, extraction. | 8,192 tokens | $0.01 | $0.04 | Starts when called |
Qwen3.5 0.8Bqwen3.5:0.8b | Tiny and very fast: classification, extraction, short replies. | 8,192 tokens | $0.01 | $0.04 | Starts when called |
Qwen3.5 2Bqwen3.5:2b | Small and quick: chat, summaries, simple questions. | 8,192 tokens | $0.02 | $0.06 | Starts when called |
Bonsai 2 9Bbonsai-2:9b | A 9B model in 3 GB: good everyday answers on a small card. | 8,192 tokens | $0.03 | $0.10 | Starts when called |
Qwen3.5 4Bqwen3.5:4b | Small and capable: chat, summaries, simple code. | 8,192 tokens | $0.03 | $0.10 | Starts when called |
Qwen3.5 9Bqwen3.5:9b | A strong all-rounder: chat, writing, code and reasoning. | 8,192 tokens | $0.05 | $0.15 | Starts when called |
Ornith 1.5 9Bornith-1.5:9b | Made for agents and code: tool calls and tasks in several steps. | 8,192 tokens | $0.05 | $0.15 | Starts when called |
Nemotron 3.5 Bolt E4Bnemotron-3.5-bolt:e4b | NVIDIA’s newest small model: quick chat, reasoning and tool use. | 8,192 tokens | $0.04 | $0.12 | Starts when called |
Bonsai 2 27Bbonsai-2:27b | A 27B model in 8 GB: a big model’s answers on a mid-range card. | 8,192 tokens | $0.08 | $0.25 | Starts when called |
gpt-oss 20Bgpt-oss:20b | OpenAI’s open-weight reasoning model. | 8,192 tokens | $0.05 | $0.20 | Starts when called |
Qwen3.8 27Bqwen3.8:27b | The strongest here: hard questions, long code, research and agents. | 8,192 tokens | $0.15 | $0.60 | Starts when called |
Nemotron 3.5 Lightning 30B-A3Bnemotron-3.5-lightning:30b | NVIDIA’s mixture of experts: a big model’s answers at a small one’s speed, on a 32 GB card. | 8,192 tokens | $0.08 | $0.30 | Starts when called |
Prices are per million tokens, counted with each model's own tokenizer. Answering now: machines are serving it this minute. Starts when called: the first call starts it on a machine — a few minutes — and says when to come back.
Call one
Point any OpenAI SDK at the address below with your account key (gpuapi_… from /developers) and ask for a model by its name.
from openai import OpenAI
client = OpenAI(base_url="https://rentgpupower.com/api/v1/compute",
api_key="gpuapi_...")
r = client.chat.completions.create(model="qwen3.5:9b",
messages=[{"role": "user", "content": "Hello!"}])
print(r.choices[0].message.content, r.usage)
curl -N https://rentgpupower.com/api/v1/compute/chat/completions \
-H "Authorization: Bearer $GPURENT_API_KEY" -H "Content-Type: application/json" \
-d '{"model":"qwen3.5:9b","stream":true,"messages":[{"role":"user","content":"Hello!"}]}'
What you pay for
Exactly the tokens
Your prompt is counted with the model's own tokenizer on our server, and so is every word of the answer that reaches you. The usage in each reply is that count — never a number the machine reported.
A failed answer costs nothing
If the machine serving an answer fails part-way, the call is not charged. If it fails before its first word, another machine answers instead. If you hang up, you pay for what was generated until then.
From credits you bought
Shared models are paid from credits you bought or earned. Free credit does not pay for them: the machines that answer are paid in cash.
Your limits hold
The most a call could cost — its prompt and max_tokens — is held against your balance, your key's spend limit and your monthly cap before it runs. Only what it used is taken.
How it grows with the calls
Every machine serving a model takes its share of the calls. When they are all busy, the model is started on one more machine; when calls stop, machines go back to their owners. A renter always comes first: a machine someone rents stops serving at once. Before any machine takes a caller's call, it must answer a check question right — and it is asked more while it serves.
Private work: a machine of your own
Your prompts and answers pass through machines that belong to other people. Do not send anything private to a shared model. For that, rent a machine of your own and run the model there — an instance answers on the same API, billed by the second.
Questions about shared models
Which API is it?
OpenAI's chat completions: POST /api/v1/compute/chat/completions, streamed or whole, with tools and JSON output as each model supports them. GET /api/v1/compute/models lists the models with their prices and state.
Why did my first call say the model is loading?
A model nobody called lately is started on a machine when you ask for it — a few minutes the first time. The answer is 503 model_loading with a Retry-After header saying when to come back; OpenAI's SDKs wait and retry on their own.
How long can a prompt be?
Up to the context shown above, prompt and answer together. An answer is as long as max_tokens, or 2,048 tokens when you do not say.
Can my machine earn by serving them?
Yes. A machine in the API pool serves shared models while nobody rents it, and its host is paid of what callers pay for the tokens it serves. Hosts can turn it off in each machine's settings.
Can one huge model run across all the machines?
Not across home computers on the internet: every word of an answer would wait for every machine in turn, so it would be far slower than one machine. What scales is many copies of a model, each on one machine, sharing the calls — that is what this is. A model too big for one card runs on one machine with enough memory, or across a cluster of machines on one fast network.
Need your own GPU instead?
Run any model you like on a machine of your own, by the second — instances — or rent several machines as one cluster.