Snowflake offers serverless fine-tuning for open-weight models

Snowflake aims to take GPU management off developers' hands

Snowflake announced the public preview of Cortex Training on November 3. The service is designed to let enterprise AI teams customize and deploy open-weight large language models using their own data, without having to build, schedule or manage GPU infrastructure, checkpoints or distributed execution.

According to the company, production AI applications generate large volumes of commercially valuable data. This can include customer requests, usage patterns and outcomes. Cortex Training connects this data to the model customization process, while Snowflake manages the GPUs, model state and distributed execution.

Snowflake said the service is built on a customizable, optimized open-source backend. Based on the company's comparative benchmarks, Cortex Training can offer costs up to 61 percent lower than other serverless services.

SFT, RL and custom training loops

Cortex Training supports supervised fine-tuning, or SFT, as well as reinforcement learning, or RL. Both methods can be used with full-parameter training or LoRA. Developers can work with open-weight model families including Qwen, GLM, Nemotron, Gemma and DeepSeek.

During SFT, the model learns from high-quality examples. For a customer service agent, this could include learning to identify issues, communicate with customers, apply escalation rules, and use tools and workflows appropriately.

RL, by contrast, shapes behavior based on feedback from measurable outcomes. Snowflake's examples include teaching a customer service agent when to use a ticketing system, when to call a billing API and when to escalate a case to a human. The goal is to improve the balance between quality, speed and cost.

The service can be integrated with frameworks including Hugging Face TRL, PrimeRL, veRL and SkyRL. Lower-level APIs also allow developers to create custom loss functions, model architectures and training loops.

From data preparation to OpenAI-compatible endpoints

Snowflake intends Cortex Training to manage the entire model lifecycle. Teams can first compare models on their own workloads based on quality, latency and cost, then select the base model best suited for customization.

On Snowflake's data platform, companies can clean, filter, enrich and label structured and unstructured data. Using Cortex AI Functions, SQL and Snowpark, they can turn this data into training-ready datasets. Cortex AI Gateway can also collect agents' traces, logs and model interactions.

Together with application outcomes and human feedback, this data can form the basis for new training datasets. Snowflake describes this as a continuous improvement cycle: based on signals from production use, models can become increasingly tailored to a company's specific tasks.

Checkpoints from customized models can be compared and then deployed through OpenAI-compatible inference endpoints. Snowflake said role-based access control covers the training data, model training, deployment and inference, while the data, model weights and optimizer state remain in Snowflake's managed environment.

What does this mean for enterprise users?

Cortex Training offers an alternative for enterprise AI teams that want to improve the performance of open-weight models using their own data but do not want to operate a dedicated GPU cluster. Snowflake said developers can focus on model logic and experimentation while the platform handles infrastructure-intensive operations such as forward and backward propagation, optimization steps, weight synchronization and generation.

According to Snowflake, ResolveAI uses the serverless APIs to train long-running tasks requiring hundreds, and in some cases more than a thousand, GPUs. Researchers at OpenHands said Cortex Training allowed them to focus on experimentation and training without giving up the frameworks they had previously used.