How QwenCloud bills fine-tuning training, including the cost formula and where to find the estimate.
Fine-tuning training is billed by the number of training tokens processed. Training runs in the region where you create the job. Fine-tuning is available only in Singapore and China North 2 (Beijing). See Feature availability by region.
When you create a fine-tuning job, the final step shows the estimated cost for the selected model and dataset size. See Create a fine-tuning job.
Training does not include inference. To call the trained model, deploy it first; billing for a deployment starts as soon as it reaches Active status, even if no requests are sent. See Deployment overview.
Cost formula
- Minimum billing unit: 1 token.
- Total training tokens: The number of tokens in your published dataset that the job trains on. Samples longer than the
max_lengthhyperparameter are discarded, so they are not billed. See Hyperparameters reference. - Number of epochs: The
n_epochshyperparameter. The default is 3. - Training unit price: Depends on the base model. Your bill shows the rate applied to each training job.
Where to check usage and cost
- Before training: The estimated cost shown on the final step of the create page.
- During and after training: The Details tab of the job shows the configuration and token consumption. See Manage fine-tuning jobs.
- Final charge: Your bill. The console estimate is for planning only.
Related charges
Training does not include inference. To call the trained model, deploy it first; billing for a deployment starts as soon as it reaches Active status, even if no requests are sent. See Deployment overview.
Next steps
- Fine-tuning overview -- Training method, training mode, and supported models
- Create a fine-tuning job -- Create a job in the console
- Manage fine-tuning jobs -- Monitor progress and publish checkpoints