Zum Inhalt

Base Models

Base Models refer to pretrained foundation models that are manually loaded into the DeepExtension environment and serve as the starting point for fine-tuning and inference. These models form the backbone of your LLM development pipeline.

Note: Adding a base model is required before you can initiate training or perform fine-tuning in DeepExtension. We classify this as a developer task since it requires access to the backend file system — typically only admin users have the necessary permissions.
See How to Add Base Models for step-by-step guidance.

Overview

On the main Base Models page, you’ll see a list of all currently loaded base models. Each base model entry includes:

  • Technical Name: A short, unique identifier used internally. This is critical for naming trained models (see Customized Models for naming rules).
  • Model Name: The full official name of the model (e.g., "Qwen-1.5-7B", "LLaMA2-13B").
  • Path: A relative path to the directory where the model is saved locally.
  • Model Card Info: Additional technical details automatically extracted from a *.json file located in the model folder during registration.

Both Technical Name and Model Name must be unique across the entire system.

Add a New Base Model

To register a new base model:

  1. Click "Add a New" on the Base Models page.

  2. Provide the following information:

    • Technical Name (Model Alias) (Must be unique and concise)
    • Model Name (Base Model Name) (Complete descriptive name)
    • Path (Model Path) (Relative path to the model folder)
    • Deployment Environment (Environment used for deploying the model)
    • Model Type (Functional types supported by the model)

Once saved, DeepExtension will attempt to load metadata and configuration details from the model directory, forming a model card that can be reused by training or inference workflows.

You are responsible for ensuring the model files are fully and correctly placed in the target path.

Notes on Deployment Template

  • The Template for Model Deployment is only required when deploying models using Ollama. If you do not plan to use Ollama, you may leave this field blank.
  • You can always update this field later by clicking "Update Deployment Template File" on the Base Models page.
  • For more information about what a Modelfile is, what the Template for Model Deployment refers to, and how to prepare one, please refer to the developer guide: How to Add Base Models

Base Model Behavior

  • Base models must be manually downloaded and stored locally before registration.
  • Online model fetching (e.g., from HuggingFace or other public hubs) is not currently supported.
  • Remote integration for auto-fetching is planned for a future update and is on our development roadmap.

DeepExtension — Ground your AI workflows on reliable, local foundation models