Skip to main content
Fireworks RFT helps you train frontier models like DeepSeek V3 and Kimi K2 to outperform closed models for your product use case, using reinforcement learning. Fireworks RFT is powerful and easy to use for developers and enterprises:
  • No infrastructure: Train frontier models without managing GPUs or RL infra
  • Production-ready: Built-in tracing, monitoring, security & one-click deploy
  • Fast iteration: From evaluator setup to deployed model in hours, not weeks
See how Genspark and Vercel used Fireworks RFT to train open models for agentic use cases, outperforming leading closed models.

​
Quickstart: Pick Your Training Approach

Single-Turn Training

⏱️ 15 minutesBest for: Testing locally, simple task trainingHow it works: Iterate on your evaluator and use it to train a small model on Fireworks.

Remote Agents

⏱️ 1-2 hoursBest for: Agents, multi-turn workflows, existing servicesHow it works: Rollouts happen in your environment. Connect via HTTP with tracing.

Secure Training (BYOB)

⏱️ 2-4 hoursBest for: Sensitive data, compliance, enterpriseHow it works: Training data never leaves your GCS/S3 bucket. Full data isolation.

​
Launch Training

Prerequisites & Validation

Requirements, validation checks, and common errors before launching

CLI (eval-protocol)

Fast, scriptable, reproducible. Perfect for automation and iteration

Web UI

Visual, guided, beginner-friendly. Great for exploring options
Already familiar with firectl? You can create RFT jobs directly.

​
RFT Concepts

How RFT Works

The RL training loop explained

Evaluators

How reward functions guide training

Environments

Local vs remote evaluation environments

Parameter Tuning

Optimize your training configuration