Fireworks AI is a platform for running, combining and training open AI models. The site describes three parts: inference, model ensembling and training. Inference serves open models through serverless endpoints or inside your own cloud, with one endpoint across 30+ regions and data residency options. Ensembling routes each turn of a task to a model chosen for accuracy and cost. Training covers supervised fine-tuning, DPO and ORPO through a managed service, or a custom training loop on dedicated trainers.
The platform is aimed at businesses building AI products or agents that want to use open models instead of closed frontier APIs, and want more control over cost and data. The site lists enterprise, startup, code assistance, conversational AI, agentic systems, search and enterprise RAG as use cases, and says it is trusted by more than 10,000 companies.
You call a model through a single endpoint and choose a mode per request, such as Standard, Priority, Fast, Batch or Spot. Serverless usage is billed per token on input, cached input and output. Cached input is billed at half price or less. Teams that need guaranteed capacity can reserve throughput with an SLA, or run Fireworks Sovereign on GPUs in their own cloud account or data center.
Serverless pricing starts at $0.05 per million input tokens, with no commitment, drawn from prepaid credits. Reserved Throughput and Fireworks Sovereign are priced by contract. Managed training starts at $0.50 per million training tokens, and the Dedicated Training API starts at $8.00 per GPU-hour. The page says the Muse Glimmer 30B SFT model is free for training through November 30, with no credit card required.
