Fireworks AI

Fireworks AI is an inference and fine-tuning platform for open large language models, built for developers and AI product teams, billed per token or GPU hour.

Fireworks AI lets a software team put an open-weight language model into production without buying or running GPUs of its own. Fireworks AI, Inc., an independent company, develops and sells it. The same account carries a project from a first API call to a tuned model on dedicated hardware. A team picks a model from the library, calls it through one API, and can later train it on its own data and serve the result at the base model's rate. Serverless Inference answers requests per token with no setup and no cold starts, in Standard and Priority modes. The Batch API takes large asynchronous jobs at a lower rate than live calls. Managed training covers LoRA and full-parameter supervised tuning and DPO, while the Serverless Training API adds reinforcement learning on a shared trainer. On-demand deployments place a base or tuned model on dedicated GPUs billed by the second, and Reservations hold capacity under a longer commitment. Across models the API offers tool calling, JSON schema and grammar output, a reasoning effort setting, prompt caching, embeddings and image or audio input where a model accepts it. The platform suits developers and product teams at software companies building coding assistants, chat products and agents. It also fits a company that wants a specialized model it controls rather than a closed one it rents. Buying starts self-serve: a team adds a payment method, buys credits, and every request draws down the balance, with Auto Reload and a monthly spend limit to keep it running. Larger customers go through sales for reserved capacity, Reserved Throughput under an SLA, Zero Data Retention, single-region data residency, SSO and audit logs. Code assistance, conversational AI and agentic systems are the main jobs it runs, along with bulk request sets that do not need an answer in real time. Vercel, Sourcegraph and Cresta are among the companies that run on it. A typical project starts with prompts in the Playground, moves to serverless calls from the team's own application, and ends on a dedicated deployment once traffic is steady. Datasets and Evaluations in the console keep training data and quality checks beside the models they shape. It fits the team whose product depends on a model it can tune, measure and keep.