Parasail is an inference cloud: a service that runs trained AI models on its fleet of GPUs and answers requests from other programs, so the customer does not buy or manage hardware. It hosts 41 or more open and frontier models, among them Z.AI's GLM-5.3, Moonshot's Kimi K3, DeepSeek V4 Flash and MiniMax models, and also customers' own fine-tuned models.
All models are called through one OpenAI-compatible API, an interface that follows the request format of OpenAI's API, so existing client code can be pointed at it. Serverless endpoints have no minimum spend and are billed per million input and output tokens. Elastic endpoints give selected models dedicated GPU performance at 1.25 times the serverless token rates, with no GPU-hour commitment. Dedicated deployments reserve GPUs such as the NVIDIA B300, and batch jobs run offline on spare capacity at lower rates. Parasail states that it applies no hidden quantization, a compression of model weights that can lower quality.
Pricing is pay per use.