Inference

Upload an ONNX model and GridPup serves it on a GPU at an https address. You get a key, the model's inputs in plain words, and calls that work as they are in Python, curl and JavaScript. Language models (Llama, Phi, Qwen, Mistral and others exported for onnxruntime-genai) answer with OpenAI's chat API, so OpenAI's libraries work by changing the address. You pay for the machine by the hour while the endpoint runs, from your GridPup balance; stop it when you do not need it.

Want an open model instead (Qwen, Mistral, gpt-oss, DeepSeek and more)? Pick one from the models catalog and deploy it on the machine you choose, no upload.

1 Model

Ready-made

No ready-made chat model (llm) yet. Bring your own below.

Your model

Choose the model folder, or drop it hereThe folder Olive or onnxruntime-genai's model builder wrote: genai_config.json, the model with its weights, the tokenizer. Up to 30 GB. Log in to upload.

2 Hardware

Choose a model to see which GPUs fit it.

3 Configuration

Endpoint name

Protected: every call needs one of the endpoint's keys (the first is shown when it is created). The machine runs only while the endpoint is started; Stop ends it.

Loading…

How it works

Check. GridPup reads the model, works out what each input is (token IDs, an image, a mask) and runs a sample on CPU. A model that cannot be served is refused with the reason.
Start. GridPup rents the cheapest machine that fits the model (Auto), or one with the GPU you chose, always with enough GPU memory, memory and disk, and installs onnxruntime with CUDA.
Ready. Once the endpoint answers a sample correctly through its address, it is Ready. Call POST /v1/infer (NumPy .npz) or /v1/infer.json with your key; a language model: POST /v1/chat/completions, streaming or not.
Stop. Stop terminates the machine; the model, keys and examples stay for the next Start.

A .onnx file up to 2 GB (vision, embeddings, classifiers, detectors: one answer per call), or a language model's folder up to 30 GB, uploaded as it is.