Fireworks AI

Fireworks AI is a platform for running and customising open-weight models in production. Its serverless API serves a library of open models billed per input, cached input and output token, with modes per request, including a half-price Background Mode for work that can wait and multi-day prompt caching for agents that revisit the same context. When a workload needs its own capacity, dedicated deployments run a model on reserved GPUs (H100, H200, B200, B300 and GB300) billed per minute, and Fireworks can also run its inference engine on GPUs in a customer's own cloud account or data centre.

Training is built into the same console. Managed training takes a dataset and a base model and runs supervised fine-tuning, preference tuning (DPO) or reinforcement fine-tuning, typically as LoRA adapters billed per training token, and a finished job shows its loss curve and can be deployed in one click. The console also covers datasets, evaluations, batch inference jobs, usage analytics and FireRouter, which routes each turn to the model that fits the task.

At its first Forge conference on 3 November 2026, Fireworks launched an ultra-fast serving tier on NVIDIA Blackwell GPUs in public preview, Background Mode, its own Ember-1 model and new training and enterprise controls. It suits engineering teams building agents and AI features on open models who want inference, fine-tuning and evaluation in one place.