Skip to main content
These examples show different ways to deploy, serve, and train models on Baseten. Choose an approach based on the model and the amount of custom code you need. If you are new to Baseten, start with Deploy your first model.

​
Engines

Deploy supported LLMs, embedding models, and other common architectures on Baseten inference engines with configuration only. You do not need Python code or a Dockerfile. See engines for architecture support, quantization options, and performance guidance.

Fast LLMs with TensorRT-LLM

Speculative decoding

Embeddings with BEI

​
Custom Docker servers

Bring an inference server such as vLLM, SGLang, or another HTTP server. Baseten runs the container, and you configure its serving behavior. See Custom Docker servers for configuration.

Deploy a Hugging Face model

Run any LLM with vLLM

Deploy LLMs with SGLang

Deploy LLMs with Ollama

Dockerized model

​
Custom Python models

Write a Truss Model class to control model loading and prediction logic. Use custom Python when an inference engine or existing server does not fit the model. See custom model code for the API.

Build and deploy a LLM

Image generation

Customize a model

​
Chains

Build multi-step workflows with models that scale independently, parallel execution, routing, and postprocessing. See Chains for the SDK.

RAG pipeline with Chains

Transcribe audio with Chains

​
Training

Train and fine-tune models on managed GPUs. Use Loops to fine-tune an LLM with LoRA or run reinforcement learning from a Python training loop. Use Training Jobs to bring your own framework and container for LLM fine-tuning or custom training. The examples below use Training Jobs:
Our training infrastructure supports popular frameworks including VERL, Megatron, and Unsloth, as well as models trained directly with Hugging Face Transformers.