Snowflake Opens Cortex Training Preview

Snowflake announced the public preview of Cortex Training on November 3, 2026. The serverless service lets teams customize and deploy open-weight large language models (LLMs) using familiar training frameworks or custom training loops, while Snowflake manages GPU provisioning, scheduling, distributed execution and checkpointing. Snowflake says its benchmarks show up to 61% lower cost than competing serverless training API offerings, including Tinker and others. The comparison used pricing published on October 22, 2026, and LoRA supervised fine-tuning workloads measured for throughput per H200 GPU.

Cortex Training supports supervised fine-tuning (SFT), reinforcement learning (RL), direct preference optimization (DPO), distillation, continued pretraining and mid-training. Teams can use LoRA or full-parameter training and frameworks including Hugging Face TRL, PrimeRL, veRL and SkyRL. Lower-level serverless APIs let advanced teams define custom loss functions, model architectures and training loops. The backend, Arctic Platform, is open source, and users can inspect its GPU code and contribute changes.

Snowflake says its platform supports a model lifecycle spanning model benchmarking, customization, checkpoint evaluation and deployment through OpenAI-compatible inference endpoints. Teams can use Snowflake’s data platform, Cortex AI Functions, SQL and Snowpark to prepare structured and unstructured enterprise data for training. Cortex AI Gateway can capture agent traces, logs and interactions; teams can combine those records with application outcomes and human feedback to prepare data for SFT, preference optimization and RL.

Role-based access controls cover training data, training, deployment and inference, while data, model weights and optimizer state stay in Snowflake’s managed environment. The service provides dedicated-GPU endpoints for streaming inference in interactive applications and agents, as well as batch inference. Snowflake says GPU-time billing means faster training reduces costs.

ResolveAI’s research lead Sean Bell said its AI SRE agents need long-horizon training jobs running on hundreds to over a thousand GPUs, and that Cortex Training’s serverless APIs let the company operate at that scale without managing infrastructure. OpenHands researchers Jeff Ma and Parinthapat Pengpun said they previously spent days to weeks managing and debugging their own training cluster; with Cortex Training, they said they could focus 100% of their time on experiments and training without changing their preferred frameworks. Snowflake researcher Boyi Liu said the company’s software-engineering agents, including Arctic-SWE, use prime-rl with Cortex Training as the backend.

Snowflake announced the public preview of Cortex Training on November 3, 2026. The serverless service lets teams customize and deploy open-weight large language models (LLMs) using familiar training frameworks or custom training loops, while Snowflake manages GPU provisioning, scheduling, distributed execution and checkpointing. Snowflake says its benchmarks show up to 61% lower cost than competing serverless training API offerings, including Tinker and others. The comparison used pricing published on October 22, 2026, and LoRA supervised fine-tuning workloads measured for throughput per H200 GPU.

Cortex Training supports supervised fine-tuning (SFT), reinforcement learning (RL), direct preference optimization (DPO), distillation, continued pretraining and mid-training. Teams can use LoRA or full-parameter training and frameworks including Hugging Face TRL, PrimeRL, veRL and SkyRL. Lower-level serverless APIs let advanced teams define custom loss functions, model architectures and training loops. The backend, Arctic Platform, is open source, and users can inspect its GPU code and contribute changes.

Snowflake says its platform supports a model lifecycle spanning model benchmarking, customization, checkpoint evaluation and deployment through OpenAI-compatible inference endpoints. Teams can use Snowflake’s data platform, Cortex AI Functions, SQL and Snowpark to prepare structured and unstructured enterprise data for training. Cortex AI Gateway can capture agent traces, logs and interactions; teams can combine those records with application outcomes and human feedback to prepare data for SFT, preference optimization and RL.

Role-based access controls cover training data, training, deployment and inference, while data, model weights and optimizer state stay in Snowflake’s managed environment. The service provides dedicated-GPU endpoints for streaming inference in interactive applications and agents, as well as batch inference. Snowflake says GPU-time billing means faster training reduces costs.

ResolveAI’s research lead Sean Bell said its AI SRE agents need long-horizon training jobs running on hundreds to over a thousand GPUs, and that Cortex Training’s serverless APIs let the company operate at that scale without managing infrastructure. OpenHands researchers Jeff Ma and Parinthapat Pengpun said they previously spent days to weeks managing and debugging their own training cluster; with Cortex Training, they said they could focus 100% of their time on experiments and training without changing their preferred frameworks. Snowflake researcher Boyi Liu said the company’s software-engineering agents, including Arctic-SWE, use prime-rl with Cortex Training as the backend.