Managed Training Overview

Train models with Fireworks-managed infrastructure — no custom code required.

Give Fireworks your data and configuration. The platform handles scheduling, training, checkpointing, and model output. Training data uses the OpenAI-compatible chat completion format, so existing OpenAI SFT datasets work with no conversion required.

How to launch managed training

These interfaces create the same underlying managed jobs:

Interface	Use when
Fireworks UI	You want guided configuration and visual monitoring
CLI or API	You want scripted and reproducible job operations
Your agent	You want help configuring, running, and troubleshooting training

The Fireworks CLI is called firectl. Install the training skill to use your agent, or continue with the method-specific managed guides below. For custom Python training loops, start with the Training API overview.

Methods

Supervised Fine-Tuning

Train text and vision models with labeled examples of desired outputs

Preference Optimization (DPO / ORPO)

Train on preferred and non-preferred response pairs using DPO or ORPO

Reinforcement Fine-Tuning

Train models using custom reward functions for complex reasoning tasks

Supported base models

Fireworks supports training for major open source model families, including DeepSeek, Qwen, Kimi, Gemma, GLM, and Llama. Eligibility is decided per model and per method: a model can support SFT without supporting DPO or RFT.

Models is the live per-model matrix: the surfaces and methods each base model is enabled for, the training shapes behind it, and each shape’s maximum context length. Check it before creating a job, and set the job context from a shape that supports the method you picked, using firectl sftj create, firectl dpo-job create, or the corresponding RFT command.

Custom models uploaded by users are not automatically tunable. To use managed training with an uploaded custom base model, the model must have a corresponding Hugging Face URL. Fireworks uses that URL to infer the training renderer and locate compatible training shapes. A custom model is supported only when Fireworks can resolve both a supported renderer and at least one compatible training shape. After the Hugging Face URL is set, tunability is refreshed by a background operation that runs about every 30 minutes, so the model may take up to 30 minutes to show as Tunable: true. We are working to make this refresh faster.

To browse the broader catalog (including non-tunable inference models), visit the Model Library for text models or vision models.

Tuning modes and context length

Managed training runs Low-Rank Adaptation (LoRA) only. It supports the full context lengths exposed by the available training shapes, matching the same long-context capabilities used by cookbook recipes.

LoRA gives you efficient adapter training and flexible deployment, including multiple LoRAs on a single base model deployment. For full-parameter tuning, use the Training API.

Deprecation notice: The deployedModel request key for routing to LoRA addons is deprecated and will not be supported for any new deployments. Please migrate to the model field with the <model_name>#<deployment_name> format described in Routing requests to LoRA addons.