Cloud & Architecture · Paid
Fireworks Training API
The Fireworks Training API is Fireworks AI's service for fine-tuning and post-training open-weight models with your own Python training loop. Your script runs on your own machine: it loads the data, builds batches, computes losses or rewards and decides each step, while Fireworks runs the forward passes, backward passes and optimizer steps on remote GPUs it manages. Trained checkpoints are promoted to models in your Fireworks account and can be deployed for inference.
Who it is for: developers and ML teams who want supervised fine-tuning (SFT), preference tuning (DPO, ORPO), reinforcement learning, distillation or custom training loops on open models without running their own GPU cluster. Since November 2026 it is also the replacement for Fireworks' managed SFT and DPO jobs, which no longer accept new jobs.
Key features: - Two infrastructure options. Serverless Training attaches to a shared, pooled trainer for LoRA work on a list of supported models, with no provisioning and per-token billing. Dedicated Training provisions trainer and deployment resources for broader model support, full-parameter training and explicit checkpoint, resume and deployment control, billed per GPU hour while the trainer runs. - Cookbook recipes. Fireworks publishes ready-made SFT, DPO, GRPO-style and experimental async RL loops in its public cookbook repository and as the fireworks-training-cookbook package on PyPI, plus a Python SDK (fireworks-ai) for full control. - Data stays with you. The API reads training data from a local JSONL file, a directory of shards or a Google Cloud Storage path; it does not create a dataset resource on Fireworks, and Fireworks describes the Training SDK as zero data retention by default. - Coding-agent support. Fireworks provides a training skill and copy-paste prompts so tools like Claude Code, Cursor or Codex can configure and run jobs. - Migration guides from Fireworks managed training and from Thinking Machines' Tinker API.
Pricing type: paid. Serverless Training is billed per token (prefill, sampling and training tokens listed separately by model); Dedicated Training uses the same GPU rate card as Fireworks On Demand deployments, billed per GPU hour. Check Fireworks' pricing page for current rates.
Things to know: the training loop runs in your process, so if your machine or script stops, training stops; long runs belong on a machine that stays on. A dedicated trainer stops after 60 minutes without activity and keeps billing until then.
Alternatives: Thinking Machines Tinker, Together AI fine-tuning, OpenAI fine-tuning (for OpenAI models), Modal or RunPod with open-source trainers such as Hugging Face TRL, or self-managed GPUs.
Official site: https://docs.fireworks.ai/fine-tuning/training-api/introduction
