Parasail

Parasail is an LLM inference cloud that serves open models through one OpenAI-compatible API for AI startups, paid per token with no minimums.

Parasail lets a developer call open-weight models such as DeepSeek, Qwen, Kimi, GLM, Llama and gpt-oss from the same client code, on GPUs the team never has to provision. Requests follow the OpenAI format, so the OpenAI SDKs work as clients. The service is aimed at AI-native startups that need inference to stay reliable as usage grows. Parasail, Inc., an independent company, owns and operates it. Serverless, Elastic, Dedicated and Batch share one pricing page and one API. Serverless endpoints give shared per-token access with no minimums or rate limits, and Elastic endpoints put selected models on dedicated GPUs without an idle bill. Dedicated deployments reserve GPUs for one customer with a negotiated latency SLA, while Batch runs offline jobs of millions of requests on spare capacity. The API covers tool calling, structured output held to a supplied JSON schema, embeddings and image input on vision-language models such as Qwen-VL. Reasoning depth is set per model, through reasoning effort on GPT-OSS and a thinking toggle on Qwen. Developers and startups that build on open models, and want frontier releases available on day zero, are the core audience. Teams leaving a closed model API can keep their OpenAI client libraries. Groups with heavy offline workloads, such as model evaluations, send them to Batch for the lowest per-token rate. Self-serve use is paid as you go by the token, charged to a card after use. Larger companies that need reserved capacity and dedicated support move to an enterprise contract through sales. Parasail runs the model behind chat and text features, agents that call a team's own tools, and retrieval pipelines that pair its embeddings with a generation model. Coding agents such as Cline and desktop assistants such as AnythingLLM can use it as their model provider. A model tuned elsewhere can be served on a Dedicated deployment, including models kept in private Hugging Face repositories. On Serverless and Dedicated, personal data sent as input is not stored or logged beyond what delivering the output needs. It fits startups that want open models in production without running an inference fleet of their own.