Snowflake puts Cortex Training into public preview

Snowflake has put Cortex Training into public preview, offering a serverless service for customizing and deploying open-weight models through supervised fine-tuning, reinforcement learning and custom training loops. The service expands Snowflake’s AI stack into post-training infrastructure while leaving developers in control of their training code.

The preview documentation says access requires enablement through a Snowflake account team and is limited to commercial AWS regions. Snowflake warns that model availability, APIs and specifications may change before general availability.

Developer code, managed GPU operations

Cortex Training splits a training job between a developer-controlled workspace and Snowflake-managed compute. The workspace can run on a laptop or CPU-only virtual machine and keeps the training script, orchestration logic, data loading, batch preparation and other non-GPU work, including rollouts. Low-level APIs let teams build custom trainers and orchestrate pre-training, mid-training, supervised fine-tuning and reinforcement-learning pipelines.

Snowflake’s managed side holds model weights and optimizer state. It performs forward and backward passes, optimizer steps, weight synchronization and generation, while also handling distributed jobs, GPU scheduling, checkpointing and artifact management. Results such as training loss are returned to the client. The division gives developers control of the training loop without requiring them to operate the GPU cluster beneath it. SageMaker AI training sessions for custom model loops provide context for a comparable managed-compute approach.

The service supports LoRA, which trains smaller adapter parameters, as well as full-parameter training. Developers can use a Python client, low-level training-loop APIs or supported open-source framework integrations. Snowflake cautions that those integrations are at different stages of upstream acceptance and directs customers to their account teams for current status and working repositories.

According to Snowflake’s release notes, users can save server-side checkpoints, resume training from them and run batch inference against a checkpoint. A base model or tuned checkpoint can also be deployed to an OpenAI-compatible endpoint that supports streaming chat completions. The documented model catalog is subject to the preview’s change warning.

Billing and performance claims

On-demand use is billed in AI Credits for assigned GPU capacity, metered by the second with no minimum. Billing begins when Snowflake assigns GPU capacity and stops when that capacity is released.

Snowflake claims selected LoRA supervised-fine-tuning workloads cost up to 61% less than competing serverless training APIs. The company says it used competitor prices published on October 22, measured throughput per H200 GPU and calculated cost per million tokens. Snowflake’s footnote describes the pricing date, H200 throughput measurement and price-per-million-token formula, but the opened sources do not provide an independent reproduction, so the comparison remains a company benchmark claim.

Snowflake also says job multiplexing raises GPU utilization and yields more usable training or inference hours per dollar. No independent utilization or cost study was located.

Researchers from OpenHands, quoted by Snowflake, said the service let them focus on experiments and training without changing their preferred frameworks. Resolve AI’s research lead said its long-horizon jobs span hundreds to more than 1,000 GPUs and that Cortex Training removes cluster-management work. Those statements are customer testimonials carried by Snowflake, not independently measured outcomes.

Snowflake has released the Arctic Platform backend under the Apache 2.0 license, while a separate Cortex Training repository contains the Python SDK, command-line interface, model catalog and recipes. Snowflake has not announced a general-availability date or documented availability beyond commercial AWS regions.