Predibase offers two deployment categories:
  • Private Deployments - Dedicated resources with guaranteed availability, recommended for production use
  • Shared Endpoints - Pre-deployed models for quick experimentation and development
You can query these deployments using either the standard Predibase method or the OpenAI-compatible method.

​
Predibase Method

​
Python SDK

First, install the Predibase Python SDK: 
pip install -U predibase
Then initialize the client and start generating: 
from predibase import Predibase

# Initialize client
pb = Predibase(api_token="<PREDIBASE_API_TOKEN>")

# Get deployment client
client = pb.deployments.client("my-deployment")  # or use a shared endpoint like "qwen3-8b"

# Generate text
response = client.generate(
    "What is machine learning?",
    max_new_tokens=100,
    temperature=0.7
)
print(response.generated_text)

# Stream responses
for response in client.generate_stream(
    "Write a story about a robot learning to paint.",
    max_new_tokens=200
):
    print(response.token.text, end="", flush=True)
For vision language models, you must use the chat completions method in order to provide images in the input.

​
REST API

You can also use our REST API directly: 
# Set your credentials
export PREDIBASE_API_TOKEN="<YOUR TOKEN>"
export PREDIBASE_TENANT_ID="<YOUR TENANT ID>"
export PREDIBASE_MODEL="my-deployment"  # or a shared endpoint

# Generate text
curl -X POST https://serving.app.predibase.com/$PREDIBASE_TENANT_ID/deployments/v2/llms/$PREDIBASE_MODEL/generate \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $PREDIBASE_API_TOKEN" \
  -d '{
    "inputs": "What is machine learning?",
    "parameters": {
        "max_new_tokens": 100,
        "temperature": 0.7
    }
  }'

​
OpenAI-Compatible Method

Predibase supports OpenAI Chat Completions v1 compatible endpoints that make it easy to migrate from OpenAI to Predibase.

​
Python SDK

Use the OpenAI Python SDK with Predibase’s endpoints: 
from openai import OpenAI

# Initialize client
api_token = "<PREDIBASE_API_TOKEN>"
tenant_id = "<PREDIBASE_TENANT_ID>"
model_name = "<DEPLOYMENT_NAME>"  # Ex. "qwen3-8b"
adapter = "<ADAPTER_REPO_NAME>/<VERSION_NUMBER>"  # Ex. "adapter-repo/1" (optional)
base_url = f"https://serving.app.predibase.com/{tenant_id}/deployments/v2/llms/{model_name}/v1"

client = OpenAI(
    api_key=api_token,
    base_url=base_url
)

# Chat completion
completion = client.chat.completions.create(
    model=adapter,  # Use empty string "" for base model
    messages=[
        {"role": "user", "content": "What is machine learning?"}
    ],
    max_tokens=100
)
print(completion.choices[0].message.content)

# Stream responses
completion_stream = client.chat.completions.create(
    model=adapter,  # Use empty string "" for base model
    messages=[
        {"role": "user", "content": "Write a story about a robot learning to paint."}
    ],
    stream=True
)
response = []
for message in completion_stream:
    token = message.choices[0].delta.content
    response.append(token)
    print(token, end='')

​
REST API

Use the OpenAI-compatible REST API: 
# Set your credentials
export PREDIBASE_API_TOKEN="<YOUR TOKEN>"
export PREDIBASE_ENDPOINT="https://serving.app.predibase.com/<TENANT_ID>/deployments/v2/llms/<MODEL_NAME>"

# Chat completion
curl -i $PREDIBASE_ENDPOINT/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $PREDIBASE_API_TOKEN" \
  -d '{
    "model": "",
    "messages": [
      {"role": "user", "content": "What is machine learning?"}
    ],
    "max_tokens": 100
  }'

​
Function Calling

Function calling allows models to interact with external tools and APIs in a structured way. To do function calling with Predibase deployments and/or adapters, define your functions and include them in your requests with the OpenAI Chat Completions v1 SDK method: 
from openai import OpenAI

# Initialize client
api_token = "<PREDIBASE_API_TOKEN>"
tenant_id = "<PREDIBASE_TENANT_ID>"
model_name = "<DEPLOYMENT_NAME>"  # Ex. "qwen3-8b"
adapter = "<ADAPTER_REPO_NAME>/<VERSION_NUMBER>"  # Ex. "adapter-repo/1" (optional)
base_url = f"https://serving.app.predibase.com/{tenant_id}/deployments/v2/llms/{model_name}/v1"

client = OpenAI(
    api_key=api_token,
    base_url=base_url
)

# Define functions
tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "parameters": {
                "type": "object",
                "properties": {
                    "location": {"type": "string"}
                }
            }
        }
    }
]

# Make request with functions
completion = client.chat.completions.create(
    model=adapter,
    messages=[
        {"role": "user", "content": "What's the weather like in San Francisco?"}
    ],
    max_tokens=100,
    tools=tools
)

print("Completion result:", completion.choices[0].message.content)

​
Structured Output

Predibase endpoints allow you to enforce that responses contain only valid JSON and adhere to a provided schema. The schema can be provided either using JSON Schema (REST, Python) or Pydantic (Python).

​
Using Pydantic (Python SDK)

from pydantic import BaseModel, constr
from predibase import Predibase
import json

# Initialize Predibase client
pb = Predibase(api_token="<PREDIBASE_API_TOKEN>")

# Define a schema for the response
class Character(BaseModel):
    name: constr(max_length=10)
    age: int
    strength: int

# Get a handle to the base LLM deployment
client = pb.deployments.client("qwen3-8b")

# Generate a response that adheres to the schema
response = client.generate(
    "Generate a new character for my awesome game. Strength 1-10.",
    response_format={
        "type": "json_object",
        "schema": Character.model_json_schema(),
    },
    max_new_tokens=128,
)

# Load the response as JSON and init an object of the desired schema
response_json = json.loads(response.generated_text)
my_character = Character(**response_json)

​
Using JSON Schema (REST API)

# Export environment variables
export PREDIBASE_API_TOKEN="<YOUR_TOKEN_HERE>" # Settings > My Profile > Generate API Token
export PREDIBASE_TENANT_ID="<YOUR_TENANT_ID>" # Settings > My Profile > Overview > Tenant ID
export PREDIBASE_DEPLOYMENT="qwen3-8b"

# Query the LLM deployment
curl https://serving.app.predibase.com/$PREDIBASE_TENANT_ID/deployments/v2/llms/$PREDIBASE_DEPLOYMENT/generate \
    -H "Content-Type: application/json" \
    -X POST \
    -d '{
        "inputs": "Generate a new character for my awesome game. Strength 1-10.",
        "parameters": {
            "response_format": {
                "type": "json_object",
                "schema": {
                    "properties": {
                        "name": {"maxLength": 10, "title": "Name", "type": "string"},
                        "age": {"title": "Age", "type": "integer"},
                        "strength": {"title": "Strength", "type": "integer"}
                    },
                    "required": ["name", "age", "strength"],
                    "title": "Character",
                    "type": "object"
                }
            },
            "max_new_tokens": 128
        }
    }' \
    -H "Authorization: Bearer ${PREDIBASE_API_TOKEN}"

​
Complex Schemas

You can define more complex schemas with nested objects and arrays: 
from pydantic import BaseModel, Field
from typing import List

class Skill(BaseModel):
    name: str
    level: int = Field(ge=1, le=100)
    description: str

class Character(BaseModel):
    name: str
    age: int
    skills: List[Skill]
    inventory: List[str]
    stats: dict[str, float]

response = client.generate(
    "Create a detailed RPG character with skills and inventory",
    response_format={
        "type": "json_object",
        "schema": Character.model_json_schema(),
    }
)