> ## Documentation Index
> Fetch the complete documentation index at: https://fireworks.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Managed Training Overview

> Fine-tune open-weight models on Fireworks-managed infrastructure with no custom training code, using supervised fine-tuning or preference optimization.

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](https://fireworks.ai/docs/fine-tuning/agent/use-with-coding-agents.md) to use your agent, or continue with the method-specific managed guides below. For custom Python training loops, start with the [Training API overview](https://fireworks.ai/docs/fine-tuning/training-api/introduction.md).

## Training metrics

Per-step loss and learning-rate artifacts are retained for 7 days after they are written. After that, the job page explains that the chart is no longer available. For longer-term metric history, configure [Weights & Biases](https://fireworks.ai/docs/ecosystem/integrations/wandb) before launching the job.

## Methods

> [card] [Supervised Fine-Tuning](https://fireworks.ai/docs/fine-tuning/fine-tuning-models.md)
>
> Train text and vision models with labeled examples of desired outputs

> [card] [Preference Optimization (DPO / ORPO)](https://fireworks.ai/docs/fine-tuning/dpo-fine-tuning.md)
>
> Train on preferred and non-preferred response pairs using DPO or ORPO

> [card] [Reinforcement Learning](https://fireworks.ai/docs/fine-tuning/training-api/cookbook/rl.md)
>
> Train against a reward function on the Training API. Managed RFT has been removed.

## 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.

[**Models**](https://fireworks.ai/docs/fine-tuning/models.md) 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` or `firectl dpo-job create`.

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](https://app.fireworks.ai/models?filter=LLM\&tunable=true) or [vision models](https://app.fireworks.ai/models?filter=vision\&tunable=true).

## Tuning modes and context length

Managed training runs **[Low-Rank Adaptation (LoRA)](https://arxiv.org/abs/2106.09685)** by default. 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](https://fireworks.ai/docs/fine-tuning/deploying-loras.md#multi-lora-deployment) on a single base model deployment.

Select models also support full-parameter SFT and DPO in managed training. To use it, [set the LoRA rank to `0`](https://fireworks.ai/docs/fine-tuning/fine-tuning-models#additional-managed-training-job-settings). If the model doesn't support it, the job is rejected with "model is not tunable for supervised full-parameter fine-tuning". Full-parameter training produces a full model rather than an adapter, so it can't be served as a multi-LoRA addon. For most full-parameter work, we recommend the [Training API](https://fireworks.ai/docs/fine-tuning/training-api/introduction.md), which has ready-made recipes and works with the [training skill](https://fireworks.ai/docs/fine-tuning/agent/use-with-coding-agents.md).

> [!WARNING]
>
> **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](https://fireworks.ai/docs/fine-tuning/deploying-loras.md#multi-lora-deployment).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.