Fireworks AI

Fireworks AI is an infrastructure platform for building with open AI models. It covers both training and inference, with the aim of letting companies turn leading open models and their own data into specialised models they control. Fireworks says it was founded by the creators of PyTorch, and its CEO is Lin Qiao.

Training

Fireworks offers several ways to train or customise a model, depending on how mature your workload is. In a guided path, you describe the task, review the plan and cost, approve the run and receive a trained model. In a configuration-led approach, you choose the model, data and method, and Fireworks handles scheduling, training and the handoff to production. For advanced teams, you can write your own training logic, including your own loss function, trainer and reinforcement learning loop, running on Fireworks GPUs with rollout serving and weight synchronisation. Each checkpoint can be deployed to production in seconds.

Inference

Fireworks serves the latest open models, or your own trained versions, from development scale up to very large production workloads.

Fireworks Nexus

Nexus is a newer product described as a drop-in replacement for closed-model APIs. It routes each task to the best open or closed model, and Fireworks claims it can cut AI coding spend by 50 to 75%. That is the company's own estimate.

Pricing

Serverless inference uses per-token pricing with pre-paid, usage-based billing, zero setup and no cold starts. Training lets you customise open models with your own data, with its own price list. On-demand deployments are billed per GPU second for faster speeds, higher rate limits and lower costs at scale. Enterprise deployments with custom terms are available through sales.

Who it suits

It suits AI engineering teams that want to run open models in production, fine-tune them on their own data, or reduce reliance on closed-model APIs.

Things to consider

Per-token and per-GPU-second billing can rise quickly under heavy load, so set budgets and monitor usage. Check each open model's licence for commercial use. Before sending sensitive data for training, review data retention and security terms.

Teams can start on shared serverless endpoints while prototyping and later move to dedicated deployments or custom-trained models as traffic and quality requirements grow.