Models

Which base models you can train on Fireworks and the surfaces each one is available on.

Managed fine-tuning, the Training API, and serverless training all draw from the same base model catalog, but availability is decided per model: managed jobs by method (SFT, DPO, RFT), Training API jobs by parameter mode (LoRA or full-parameter).

Model availability

Pick a model to see the surfaces and methods it is enabled for, plus any training shapes that back it. Switch to All models for the full matrix.

Vision and multimodal support

Vision support is model- and surface-specific. Use the catalog above to confirm that the selected VLM has a compatible managed method or Training API shape before preparing data.

• Managed VLM SFT dataset schema and launch flow: Supervised Fine-Tuning: Vision
• Training API VLM loops: start from a VLM-compatible shape and the same cookbook SFT, DPO, or RL recipe used for text, replacing the text tokenizer with the model processor
• Inference request formats after deployment: Vision-language models

Next steps

Managed Fine-Tuning

Hand Fireworks your data and let the platform run the job

Training API

Write your own training loop against a Tinker-compatible API

Serverless Models

Serverless Training API model catalog with per-token pricing

Training Shapes

What a shape pins and how to reference one

Dedicated Training

Provision a trainer and sampler on reserved GPU capacity

Pricing

Current rates across training and inference