Fireworks AI is an inference and training platform for open AI models, made by Fireworks AI, Inc. Inference means running a trained model to produce outputs. Fireworks runs open models such as DeepSeek, Kimi, GLM, Qwen and OpenAI gpt-oss on its own GPU servers and exposes them through an API.
A program sends a request to an API that works as a drop-in replacement for the OpenAI API and gets text, vision, audio or embedding results back. Serverless inference needs no setup: popular models are shared and billed per token. Dedicated deployments reserve GPUs for one customer, with autoscaling, and some models need them, including Qwen3 Omni, which accepts audio input and can transcribe it. Customers can also fine-tune models on their own data and serve the result. The service runs in more than 30 regions with data-residency options.
It is a hosted, pay-per-use service: serverless use is billed per million tokens from prepaid credits, dedicated deployments per GPU second, and fine-tuning per million training tokens.