Pricing

Overview

Fine-tuning pricing at Together AI is based on the total number of tokens processed during your job. This includes both training and validation processes, and varies based on the model size, fine-tuning type (Supervised Fine-tuning or DPO), and implementation method (LoRA or Full Fine-tuning).

How Pricing Works

The total cost of a fine-tuning job is calculated using:

Model size (e.g., Up to 16B, 16.1-69B, etc.)
Fine-tuning type (Supervised Fine-tuning or Direct Preference Optimization (DPO))
Implementation method (LoRA or Full Fine-tuning)
Total tokens processed = (n_epochs × n_tokens_per_training_dataset) + (n_evals × n_tokens_per_validation_dataset)

Each combination of fine-tuning type and implementation method has its own pricing. For current rates, refer to our fine-tuning pricing page.

Token Calculation

The tokenization step is part of the fine-tuning process on our API. The exact token count and final price of your job will be available after tokenization completes. You can find this information in:

Your jobs dashboard
Or by running together fine-tuning retrieve $JOB_ID in the CLI

Frequently Asked Questions

Is there a minimum price for fine-tuning?

No, there is no minimum price for fine-tuning jobs. You only pay for the tokens processed.

What happens if I cancel my job?

The final price is determined based on the tokens used up to the point of cancellation.

Example:

If you're fine-tuning Llama-3-8B with a batch size of 8 and cancel after 1000 training steps:

Training tokens: 8192 [context length] × 8 [batch size] × 1000 [steps] = 65,536,000 tokens
If your validation set has 1M tokens and ran 10 evaluation steps: + 10M tokens
Total tokens: 75,536,000
Cost: Based on the model size, fine-tuning type (SFT or DPO), and implementation method (LoRA or Full FT) chosen (check the pricing page)

How can I estimate my fine-tuning job cost?

Calculate your approximate training tokens: context_length × batch_size × steps × epochs
Add validation tokens: validation_dataset_size × evaluation_frequency
Multiply by the per-token rate for your chosen model size, fine-tuning type, and implementation method