You run it on your laptop. Now ship it.

BitsRun turns any open model you use with Ollama or Hugging Face into your own private API. It scales to zero when nobody is calling it, and you pay by the second only while it runs.

Deploy your first model See prices

From $0.86 per GPU hour, billed per second. $1 of free credit to try it.

Terminalqwen3-8b, NVIDIA L4

$ bitsrun deploy ollama:qwen3:8b

Not deployed$0.00024 per second

$0.00000

0.0 s billed

From local model to private API in three steps

Step 1
Pick the model you already run
An Ollama name like qwen3:8b, a Hugging Face repo, or your own fine-tune as GGUF or safetensors.
Step 2
Deploy it with one command
bitsrun deploy gives you a private endpoint on a GPU that fits the model. No Dockerfile, no server to manage.
Step 3
Call it from your app
Use the OpenAI or Ollama client you already have. Change the base URL, keep your code.

Pay for the minutes you use, not the month

Most apps are busy for a few hours a day. Move the slider to see what that costs compared with keeping the same GPU running all month.

BitsRun, scales to zero

$77.76 / month

3 h a day for 30 days at $0.86 per hour

Same GPU running all month

$630.72

You keep 88% of that bill.

Keep your code. Change one line.

Endpoints speak the OpenAI and Ollama APIs, with streaming. Point your existing client at your BitsRun URL and add your key.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.bitsrun.com/v1/e/qwen3-8b",
    api_key=os.environ["BITSRUN_API_KEY"],
)
reply = client.chat.completions.create(
    model="qwen3:8b",
    messages=[{"role": "user", "content": "Hello"}],
)

How billing works

Startup is never billed: The meter starts when your model is ready to answer and stops when your idle timeout ends.
No surprise bills: You buy credits up front and set a monthly cap. When the cap is reached, your endpoint scales to zero.
Your prompts stay yours: We never log prompts or outputs, and we never train on them. Each endpoint is private to your account.
Requests wait, they do not fail: If your endpoint is waking up, requests are held until it is ready instead of returning errors.

Questions, answered

Which models can I run?

Large language models in GGUF (through llama.cpp) or safetensors (through vLLM), plus embedding and reranking models. Image and audio models are next on the roadmap.

How long does a cold start take?

When a GPU is already warm, a 7 to 8B model is usually ready in 5 to 20 seconds. Starting a new GPU takes about one to two minutes. Neither is billed.

What does the idle timeout do?

It is how long your endpoint stays up after the last request, so the next request answers instantly. You choose it, from 10 seconds to 60 minutes, and you pay per second while it waits.

Can I keep an endpoint always on?

Yes. Always-on replicas have a fixed monthly price, 10% less than running the same GPU by the hour all month.

Do I need a GPU or Docker?

No. If the model runs in Ollama or loads from Hugging Face, BitsRun can serve it. You never write a Dockerfile.

Ship the model you already trust

Deploy your first model