Predibase
Predibase is a low-code AI platform designed specifically for developers. It aims to provide a fast and efficient way to train, finetune, a...
Last verified:
What is Predibase?
Predibase is a developer platform for fine-tuning and serving open-source Large Language Models (LLMs) in your cloud. Built by the creators of open-source frameworks LoRAX and Ludwig, Predibase provides an integrated platform for hosting, customizing, and querying LLMs with production-ready private serverless endpoints.
Key features include LoRAX technology that enables serving hundreds of fine-tuned models on a single GPU, dramatically reducing costs. The platform supports supervised fine-tuning, reinforcement fine-tuning, and continued pretraining. Users can deploy via UI or Python SDK, access the latest open-source models including LLMs, VLMs, and embedding models, and choose between SaaS or VPC deployment options on AWS, Azure, or GCP.
Predibase is designed for engineering teams, researchers, data scientists, and enterprises who want to productionize open-source LLMs efficiently. It's used by Fortune 500 companies and high-growth companies to fine-tune models like Llama-2, Mistral, Falcon, and Qwen for specialized AI applications while maintaining data control and security with SOC-2 compliance.
The platform enables cost-effective model serving by putting hundreds of fine-tuned models into production for the cost of serving one. Private deployments ensure no data sharing and full model ownership, while serverless infrastructure provides autoscaling and scale-to-0 capabilities.
Predibase pricing
Pricing model: Paid
Predibase offers a Free Plan and Enterprise Plan. Free Tier includes: up to 1 user, best-in-class fine-tuning with A100 GPUs, 1 private serverless deployment (no rate limits), autoscaling and scale to 0, unlimited adapters on single GPU with LoRAX, free shared serverless inference with rate limits for testing, access to all base models, data connection via file uploads, 2 concurrent training jobs, and in-app chat/email/Discord support. Free credits expire after 30 days. Enterprise Plan includes: additional seats, volume discounts, guaranteed instances, additional replicas for burst, additional private serverless deployments, guaranteed uptime SLAs, data connections via Snowflake/Databricks/S3/BigQuery, additional concurrent training jobs, dedicated Slack channel plus consulting hours. Enterprise VPC allows deployment into your own cloud (AWS/Azure/GCP) using your own cloud commitments and GPUs. Private Serverless Inference pricing: L4 $2.14/hr, A10G $2.60/hr, L40S $3.20/hr, A100 $4.80/hr. Shared Serverless Inference is free up to 1M tokens/day and 10M tokens/month. Fine-tuning costs vary by dataset size, base model size, and training method (LoRA, Turbo LoRA, or Turbo).
Predibase pros
- Serves hundreds of fine-tuned models on a single GPU with LoRAX
- Free tier includes up to 1 user with A100 GPUs for fine-tuning
- 1 private serverless deployment with no rate limits in free tier
- Autoscaling and scale to 0 for cost optimization
- Unlimited adapters served on single GPU deployment
- Free shared serverless inference up to 1M tokens/day for testing
- Access to all available base models including latest LLMs
- Python SDK and UI for deploying, fine-tuning, and prompting
- Deploy in your own cloud (AWS, Azure, GCP) with VPC option
- 2 concurrent training jobs included in free tier
- Data never used to train other models without explicit permission
- In-app chat, email, and Discord support included
- Fine-tuned models can beat GPT-4 performance
- Reduce inference costs by up to 80% compared to dedicated deployments
- SOC-2 compliance with enterprise-grade security
- Supports supervised fine-tuning, reinforcement fine-tuning, continued pretraining
- No upfront commit required for Free tier
Predibase cons
- Free credits expire after 30 days
- Developer tier removed as of July 2025, must upgrade to Enterprise after credits
- Enterprise plan requires upfront commit
- H100 and H200 GPUs only available for Enterprise customers
- Multi H100 or H200 deployments are Enterprise-only
- Fine-tuning jobs may queue for long time during peak times
- Training time varies from minutes to hours depending on task and dataset size
- Free tier limited to 1 user only
- Only 2 concurrent training jobs in free tier
- Data connections in free tier limited to file uploads only
Frequently asked questions about Predibase
How does pricing work?
As of July 2025, Predibase removed the Developer tier. After you run out of credits in the Free tier, you need to upgrade to Enterprise to continue using Predibase, which has an upfront commit. Pricing is consumption-based for the SaaS tier, billed by the second. Private Serverless Inference is priced per hour by hardware (L4 $2.14/hr to A100 $4.80/hr). Shared Serverless Inference is free up to 1M tokens/day and 10M tokens/month.
How can I upgrade to Enterprise?
To upgrade to Enterprise, reach out to [email protected] and the team will be in touch to discuss upgrading from the Free tier. This is necessary when you run out of free credits.
My data is very sensitive. Can I still use Predibase?
Yes, Predibase offers VPC deployment where you can deploy the Predibase Data Plane into a Virtual Private Cloud on any major cloud provider (AWS, GCP, Azure). This ensures your data never leaves your cloud. Contact [email protected] to learn more about VPC options.
Will my data be used to train other models?
Your data will never be used to train models unless you explicitly grant permission to do so. If you wish to remove any shared data from the platform, reach out to [email protected]. Additional details are available in the Privacy Policy.
Why has my fine-tuning job been queued for a long time?
Once you submit a fine-tuning job, the model remains queued until available resources are available to finish training. Queue times may increase during peak times when compute resources are in high demand.
How long does model training take?
Once training begins, model training varies and can take anywhere from a few minutes to a few hours depending on the task type, model size, dataset size, and number of epochs configured.
How do I delete my account?
You can delete your account by going to Settings > My Profile > Delete Account. If you delete your account, your datasets and adapters will continue to exist in the Team. To close your Team entirely, contact [email protected].
What models are supported for fine-tuning?
Officially supported models include Solar Mini, Solar Pro Preview, Llama 3.2 (1B, 3B), Llama 3.1 8B, Mistral-7b variants, Mistral Nemo 12B, Mixtral-8x7B-Instruct-v0.1, Codellama 13B/70B Instruct, Zephyr 7B Beta, Gemma 2 (9B, 27B), Phi 3.5 Mini Instruct, Phi 3 4k Instruct, Qwen 2.5 (1.5B, 7B, 14B, 32B), Qwen 2 (1.5B), and any OSS model from Huggingface on a best effort basis.
What is LoRAX and how does it reduce costs?
LoRAX (LoRA Exchange) is Predibase's framework that enables serving hundreds of fine-tuned models on a single GPU through dynamic adapter loading. This dramatically reduces costs by allowing unlimited fine-tuned adapters to be served on a single GPU deployment without additional overhead, compared to dedicated deployments where each model needs its own GPU.
What support options are available?
Free tier includes in-app chat, email, and Discord support. Enterprise customers get a dedicated Slack channel plus consulting hours with experts. For any errors or questions, you can reach out via in-app chat or email [email protected].