Skip to main content
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: Install the training skill or continue with the method-specific managed guides below. For custom Python training loops, start with the Training API overview.

​
Methods

Supervised Fine Tuning - Text

Train text models with labeled examples of desired outputs

Supervised Fine Tuning - Vision

Train vision-language models with image and text pairs

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

​
Free Reinforcement Fine-Tuning

Reinforcement Fine-Tuning (RFT) is free for models under 16B parameters. When creating an RFT job in the UI, filter for free tuning models in the model selection area on the fine-tuning creation page. If kicking off jobs from the terminal, you can find the model ID from the Model Library. Note: SFT and DPO jobs are billed per training token for all model sizes—see the pricing page for details.
When creating a Reinforcement Fine-Tuning job in the UI, look for the “Free tuning” filter in the model selection area: Free tuning filter in the model selection area For SFT and DPO pricing, see the pricing page.

​
Supported base models

Fireworks supports fine-tuning for major open source model families, including DeepSeek, Qwen, Kimi, Gemma, GLM, and Llama. Eligibility is method and parameter-mode specific: a model can support SFT without supporting DPO, RFT, LoRA, or full-parameter training on the same shapes. Use the live Training Shapes method-support matrix before creating a job. Custom models uploaded by users are not automatically tunable. To use managed fine-tuning 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. The table below is generated from the live training shape registry. The “Max supported context length” is a catalog-level model maximum. For a job, use the maximum context of a shape that supports the selected method and tuning mode. Set it with firectl sftj create, firectl dpo-job create, or the corresponding RFT command. 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 fine-tuning supports both Low-Rank Adaptation (LoRA) and full-parameter tuning, depending on the model, method, and selected training shape. It also supports the full context lengths exposed by the available training shapes, matching the same long-context capabilities used by cookbook recipes. Choose LoRA when you want efficient adapter training and flexible deployment, including multiple LoRAs on a single base model deployment. Choose full-parameter tuning when you need to update all model weights for difficult reasoning, alignment, or domain adaptation tasks.
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.