Predibase supports two categories of language models for deployment:
  1. Officially Supported LLMs - These are models we have first-class support, meaning they have been verified and are ensured to work well. They are also available as Shared Endpoints for SaaS customers.
  2. Custom Base Models - Predibase offers best-effort support for deploying custom base models from Huggingface.

​
Quick Start

First, install the Predibase Python SDK: 
pip install -U predibase

​
Using Shared Endpoints

Get started quickly with our pre-deployed shared endpoints: 
from predibase import Predibase

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

# Use the Mistral shared endpoint - no deployment needed
client = pb.deployments.client("llama-3-2-3b-instruct")
response = client.generate("What is machine learning?")
print(response.generated_text)

​
Creating Private Deployments

For production use cases, create your own dedicated deployment: 
from predibase import Predibase, DeploymentConfig

pb = Predibase(api_token="<PREDIBASE_API_TOKEN>")

# Create a production deployment with Llama 3
deployment = pb.deployments.create(
    name="my-llama-3-2-3b-instruct",
    config=DeploymentConfig(
        base_model="llama-3-2-3b-instruct",
        min_replicas=0,  # Scale down to 0 replicas so you don't get charged when there are no requests
        max_replicas=1,
        accelerator="l40s_48gb_100"  # Specify hardware requirements
    )
)

# Use your private deployment
client = pb.deployments.client("my-llama-3-2-3b-instruct")
response = client.generate("Explain the theory of relativity.")
print(response.generated_text)

​
Officially Supported Models

These models are fully tested, optimized, and supported by Predibase:

​
DeepSeek Models

Model NameParametersContextArchitectureLicenseGPUAlways On Shared Endpoint
deepseek-r1-distill-qwen-32b32.8B8KQwenMITA100❌
Additional DeepSeek R1 and V3 models (original or distilled) are available upon request. Contact sales@predibase.com for deploying other DeepSeek models.

​
Mistral & Mixtral Models

Model NameParametersContextArchitectureLicenseGPUAlways On Shared Endpoint
mistral-7b-instruct-v0-27B32KMistralApache 2.0A100❌
mixtral-8x7b-instruct-v0-147B8KMistral MoEApache 2.0A100❌

​
Llama 3 Models

Model NameParametersContextArchitectureLicenseGPUAlways On Shared Endpoint
llama-3-3-70b-instruct70B32KLlama-3MetaA100❌
llama-3-2-1b1B64KLlama-3MetaA100❌
llama-3-2-1b-instruct1B64KLlama-3MetaA100❌
llama-3-2-3b3B32KLlama-3MetaA100❌
llama-3-2-3b-instruct3B32KLlama-3MetaA100❌
llama-3-1-8b8B64KLlama-3MetaA100❌
llama-3-1-8b-instruct8B64KLlama-3MetaA100✅
llama-3-8b8B8KLlama-3MetaA10G+❌
llama-3-8b-instruct8B8KLlama-3MetaA10G+❌
llama-3-70b70B8KLlama-3MetaA100❌
llama-3-70b-instruct70B8KLlama-3MetaA100❌

​
Llama 2 Models

Model NameParametersContextArchitectureLicenseGPUAlways On Shared Endpoint
llama-2-7b7B4KLlama-2MetaA10G+❌
llama-2-7b-chat7B4KLlama-2MetaA10G+❌
llama-2-13b13B4KLlama-2MetaA100❌
llama-2-13b-chat13B4KLlama-2MetaA100❌
llama-2-70b70B4KLlama-2MetaA100❌
llama-2-70b-chat70B4KLlama-2MetaA100❌

​
Code Llama Models

Model NameParametersContextArchitectureLicenseGPUAlways On Shared Endpoint
codellama-7b7B4KLlama-2MetaA10G+❌
codellama-7b-instruct7B4KLlama-2MetaA10G+❌
codellama-13b-instruct13B4KLlama-2MetaA100❌
codellama-70b-instruct70B4KLlama-2MetaA100❌

​
Qwen Models

Model NameParametersContextArchitectureLicenseGPUAlways On Shared Endpoint
qwen3-8b8.19B64KQwenTongyi QianwenA100✅
qwen3-14b14.8B16KQwenTongyi QianwenA100❌
qwen3-32b32.8B16KQwenTongyi QianwenA100✅
qwen3-30b-a3b30.5B16KQwenTongyi QianwenA100❌
qwen2-5-coder-3b-instruct3.09B32KQwenTongyi QianwenA100❌
qwen2-5-coder-7b-instruct7.62B32KQwenTongyi QianwenA100❌
qwen2-5-coder-32b-instruct32.8B16KQwenTongyi QianwenA100❌
qwen2-5-1-5b1.5B64KQwenTongyi QianwenA100❌
qwen2-5-1-5b-instruct1.5B64KQwenTongyi QianwenA100❌
qwen2-5-7b7B32KQwenTongyi QianwenA100❌
qwen2-5-7b-instruct7B32KQwenTongyi QianwenA100❌
qwen2-5-14b14B32KQwenTongyi QianwenA100❌
qwen2-5-14b-instruct14B32KQwenTongyi QianwenA100❌
qwen2-5-32b32B16KQwenTongyi QianwenA100❌
qwen2-5-32b-instruct32B16KQwenTongyi QianwenA100❌
qwen2-72b72.7B32KQwenTongyi QianwenA100❌
qwen2-72b-instruct72.7B32KQwenTongyi QianwenA100❌

​
Solar Models

Model NameParametersContextArchitectureLicenseGPUAlways On Shared Endpoint
solar-1-mini-chat-24061210.7B32KLlamaCustom LicenseA100❌
solar-pro-preview-instruct-v222.1B4KSolarCustom LicenseA100❌
solar-pro-24112622.1B32KSolarCustom LicenseA100❌
solar-pro-preview-instruct (deprecated)22.1B4KSolarCustom LicenseA100❌

​
Gemma Models

Model NameParametersContextArchitectureLicenseGPUAlways On Shared Endpoint
gemma-2b2.5B8KGemmaGoogleA10G+❌
gemma-2b-instruct2.5B8KGemmaGoogleA10G+❌
gemma-7b8.5B8KGemmaGoogleA100❌
gemma-7b-instruct8.5B8KGemmaGoogleA100❌
gemma-2-9b9.24B8KGemmaGoogleA100❌
gemma-2-9b-instruct9.24B8KGemmaGoogleA100❌
gemma-2-27b27.2B8KGemmaGoogleA100❌
gemma-2-27b-instruct27.2B8KGemmaGoogleA100❌

​
Other Models

Model NameParametersContextArchitectureLicenseGPUAlways On Shared Endpoint
zephyr-7b-beta7B32KMistralMITA100❌
phi-22.7B2KPhi-2MITA10G+❌
phi-3-mini-4k-instruct3.8B4KPhi-3MITA10G+❌
phi-3-5-mini-instruct3.8B64KPhi-3MITA100❌
openhands-lm-32b-v0.132.8B16KQwenTongyi QianwenA100❌
For detailed information about how to properly prompt each model, see our Chat Templates guide.

​
Custom Base Models

Predibase allows you to deploy custom (public or private) base models from Hugging Face.

​
Model Requirements

Before deploying a custom model, verify these requirements:
  1. Architecture Compatibility
    • Uses one of the supported vLLM architectures
    • Has the “Text Generation” and “Transformer” tags
    • Does not have a “custom_code” tag
  2. Format Requirements
    • Complete model weights
    • Proper configuration files
    • Compatible tokenizer implementation
    • Correct metadata and tags
    • Clear licensing for commercial use or private model

​
Deploying Custom Models

Deploy a custom model from Hugging Face Hub: 
deployment = pb.deployments.create(
    name="my-custom-model",
    config=DeploymentConfig(
        base_model="BioMistral/BioMistral-7B",
        accelerator="l40s_48gb_100", # Required for custom models
        hf_token="<YOUR HUGGINGFACE TOKEN>", # Required for private Huggingface models
        min_replicas=0,
        max_replicas=1,
    )
)

# Create client for the custom model
client = pb.deployments.client("my-custom-model")

# Generate text
response = client.generate(
    "What are some major proteins found in the human body and what is their structure?",
    max_new_tokens=300,
    temperature=0.7
)

​
Next Steps