blog
Cortex Training Announcement
Learn about the latest capabilities available in public preview.
FEATURE IN PUBLIC PREVIEW
CortexTraining
Own your intelligence within your enterprise boundaries. Customize open-weight models on your own data to improve accuracy and cut cost on domain-specific tasks.
Near-zero GPU infrastructure management
Use serverless APIs to post-train and deploy on fully managed GPUs, so AI teams can focus on model training and customization, not infra management.
Full flexibility for custom training
Innovate faster with popular models, methods (e.g. LoRA, full parameter), frameworks and custom architectures, plus an open-source backend.
Improved cost efficiency
Built-in optimizations such as redundancy elimination, model specific fwd/bwd kernel selection and fusion can help lower training cost by up to 61%.1
Governed training and inference workflow
Train and deploy open-weight models on the same platform as your governed data
Data preparation
Make your data ready for AI training
Bring your own training data sets, or use Snowflake’s native data processing and AI capabilities to clean, deduplicate, filter and transform them right where your enterprise data lives. Capture agent traces, logs and model interactions with Cortex AI Gateway.
Post-training
Train open-weight models with complete flexibility
Build on open-weight models like Qwen, GLM or Gemma with a complete set of post-training capabilities, from supervised fine-tuning and reinforcement learning to distillation and continued pre-training, with both full-parameter and LoRA-based approaches.
Evaluate and deploy
Perform high-throughput inference on fully managed GPUs
Load post-trained models for batch inference or deploy them as REST API endpoints for evaluation, testing and agent development. During production, create OpenAI-compatible inference endpoints backed by dedicated GPUs.
Execution modes
Match GPU capacity to how your team trains
- On demand: Only pay for GPU time you need without commitment. Best suited for burst experiments, iterative fine-tuning and teams proving value before scaling up.
- Reserved capacity: Lock in dedicated GPU capacity at a predictable rate with 1-3 year terms — custom terms for teams with sustained, high-volume training workloads. Access Snowflake experts to help with your model customization journey.
Customize and deploy open-weight LLMs on proprietary data
Training frameworks
Bring existing scripts built with popular open-source SFT and RL frameworks.
Training and inference APIs
REST API endpoints support weight synchronization, forward and backward passes, optimizer steps, generation, inference and evaluation workflows.
Cortex Training Runtime
The serverless runtime executes distributed training and inference jobs with Snowflake optimizations that provide up to 61% lower cost.1
Snowflake-managed GPU layer
Pay for the GPU hours you use instead of holding idle capacity while non-GPU parts of your pipeline run.
Cortex Training
Frequently asked Questions
Answers to common questions about fine-tuning open-weight models with Cortex Training.
Why should I consider open-weight models instead of frontier models?
As access to frontier models becomes commoditized and costs continue to rise, companies are increasingly adopting open-weight foundation models such as Qwen, GLM, DeepSeek and Gemma for greater performance, control and cost efficiency.
Why do I need to post-train open-weight models?
Open-weight foundation models are generic out of the box and need to be customized to a company's specific domain, data and budgetary needs.
Does using open-weight models help with AI sovereignty?
Enterprises are looking to have full self-determination and control over their own artificial intelligence tools, data and infrastructure, resulting in greater interest and adoption in open-weight models.
When should I fine-tune an open-weight model instead of using RAG or prompting?
Fine-tune when prompting and retrieval stop improving accuracy, when you need consistent output format or behavior, or when a smaller tuned model can replace a larger, costlier one. Retrieval is still the better fit when answers depend on information that changes often.
What's the difference between LoRA and full-parameter fine-tuning?
LoRA trains small adapter layers on top of frozen base weights, so it needs less GPU memory and runs faster. Full-parameter fine-tuning updates every weight, which costs more but can lift accuracy further on demanding tasks. Cortex Training supports both.
Which open-weight models can I post-train with Cortex Training?
Cortex Training supports a curated set of open-weight models, including the Qwen3 family. See the supported models list in the documentation for current availability.
