Models

The Models section holds everything in your model catalogue: base models you've uploaded, fine-tuned outputs from training jobs, and checkpoints. It mirrors client.models in the SDK.

Model types

Models are grouped by source:

• Uploaded: base models you brought in from a local directory or HuggingFace.
• Fine-tuned: output models produced by a completed training job.
• Checkpoints: intermediate snapshots saved during training. See Storage for how the checkpoint limit keeps these in check.

Add a model

Click Upload model to bring in a base model, either from a HuggingFace repo ID (e.g. meta-llama/Llama-3.2-1B) or a local directory. Fleet computes a content hash first, so re-uploading the same model is instant and free.

Train a model

Click Train on any model to open the training console.

1. Method: choose full, lora, qlora, or dpo. Each shows a short description of the trade-off.
2. GPU tier: defaults to Auto, which lets Fleet pick a GPU that fits the model and is the most cost-effective for the run. The exact tier depends on live availability when the job starts. To pin a specific tier, click Choose a tier myself.
3. Dataset: pick an uploaded dataset (optional, but required for a cost estimate).
4. Hyperparameters: sensible defaults are pre-filled; expand Advanced to override any of them. See Training for the full list.

Once a dataset is selected, the console shows a live estimate for each tier: VRAM required, estimated time, and estimated cost. Tiers that can't fit the model are greyed out. Hit Start training to launch; you'll get the job ID, the assigned GPU, and the VRAM and cost estimate on the confirmation screen.

Deploy a model

Click Deploy on a ready model to spin up an inference endpoint. Pick the hardware tier and Fleet provisions a serverless endpoint that scales to zero when idle. See Deployments.

Manage models

Each model row links to its details and offers actions to download (via a signed URL) or delete it. Deleting a model frees its storage immediately.