Novita AI

Novita AI is a model API platform for developers: open-source LLMs behind an OpenAI-compatible API, plus GPU instances and agent sandboxes, billed per use.

Novita AI lets an application call a hosted language model one day and run the same workload on rented hardware the next, under one login. It answers a common problem for teams building with open models: the API, the GPUs and the runtime for agents usually come from three different suppliers. Its maker, also called Novita AI, runs it from novita.ai, with one balance and one set of documentation across every part. The model APIs, called Serverless Endpoints, host open models from DeepSeek, Qwen, Z.ai's GLM, Moonshot's Kimi, Meta's Llama, MiniMax, Google's Gemma and Mistral. Each model lists its context window, and the API supports function calling, structured outputs with a JSON schema, image input on vision-language models, prompt caching and a Batch API for asynchronous jobs. Embeddings and rerank endpoints sit beside the chat models. Dedicated Endpoints place a chosen or custom model on GPUs reserved for one customer, with autoscaling and scale to zero. Agent Sandbox provides secure runtimes where AI agents execute code, sized by vCPU and memory. GPU Instances rent single cards or whole multi-GPU servers, on demand or as cheaper spot capacity, with disks and network volumes attached. The platform suits developers and startups shipping AI features on open models without operating their own servers. It also fits teams whose code already speaks the OpenAI API, engineers who need a GPU for an afternoon, and builders of coding agents that need a secure place to execute code. Request content falls under a Zero Data Retention policy, with automated safety screening still applied. Buyers pay as they go, by the token for model calls, by the GPU-hour for endpoints and instances and by the second for sandboxes, with a sales contact for larger deployments. In practice a team might put a support chatbot on a DeepSeek or Qwen model, then send a nightly backlog of documents through the Batch API. Retrieval projects use the embeddings endpoint to index a knowledge base, and extraction jobs rely on structured outputs to return clean JSON. Developers wire the models into their editors through the integration guides for Claude Code and Cursor, and orchestration frameworks such as LangChain, Dify and LiteLLM connect to it as an OpenAI-style provider. When traffic grows, the same model moves to a dedicated endpoint, and an agent that writes code gets a sandbox to run and test it.