Snowflake said it opened a public preview of Cortex Training, a serverless offering for customizing open-weight LLMs and deploying them without provisioning GPUs.
Can you actually fine-tune a capable open model without babysitting a cluster? Snowflake's short answer is yes, if you let it run the infrastructure.
Cortex Training is the company's bet that post-training, the stage where a base model is adapted with your data to learn specific behaviors, should happen where the data already lives. Snowflake framed the problem plainly: every production AI app emits millions of traces about preferences and outcomes, but most teams still build on closed foundation models because training is hard.
The hard part is not just GPUs. You have to schedule jobs, synchronize weights, checkpoint state, and run distributed forward and backward passes without blowing up costs. Cortex Training flips that. You control the training loop and Snowflake handles GPU-intensive operations like optimizer steps and weight synchronization through serverless APIs, the company said.
That control matters for two methods Snowflake pushes. Supervised fine-tuning, or SFT, teaches a model what good looks like from high-quality examples, like how a support agent should diagnose, escalate, and call tools. Reinforcement learning, or RL, teaches a model how to make better decisions to achieve desired outcomes by optimizing against feedback signals such as resolution rate or latency, not just imitations.
The daily brief
Get the free daily brief with the day's AI news, new AI startups and funding rounds every morning.
Join 64,000 other subscribers.
Teams that have to change a workflow today are ML platform owners and researchers who currently keep a separate training stack. Cortex Training supports LoRA, a parameter-efficient technique that trains a small adapter instead of the full model, and full-parameter training, plus DPO and distillation, across families including Qwen, Gemma, DeepSeek, GLM and Nemotron. It plugs into frameworks teams already use such as Hugging Face TRL, SkyRL, veRL and PrimeRL, and exposes lower-level APIs for custom loss functions and architectures.
ResolveAI's Sean Bell, research lead at the SRE-agent startup, said his team runs long-horizon jobs across hundreds to over a thousand GPUs and uses the serverless APIs to avoid cluster management. OpenHands researchers Jeff Ma and Parinthapat Pengpun said they previously spent days to weeks debugging their own cluster for coding-agent RL and now run experiments without changing frameworks, according to Snowflake. Snowflake AI Research's Boyi Liu said the Arctic-SWE team builds on PrimeRL with Cortex Training as the backend.
The pitch is a flywheel inside Snowflake. You curate training data with Cortex AI Functions, SQL and Snowpark, capture agent traces with Cortex AI Gateway, benchmark models, train, evaluate checkpoints with batch or streaming inference, track runs in Snowflake ML Experiments, then deploy to OpenAI-compatible endpoints on dedicated GPUs. New production trajectories feed the next training run. Data, weights and optimizer state remain in Snowflake's managed environment under role-based access controls, the company said.
Cost is the sharp edge. Snowflake said benchmarks show up to 61% lower cost than competing serverless offerings, including Tinker, based on LoRA SFT throughput per H200 GPU and pricing published Oct. 22, 2026. That math depends on context length and throughput, so teams with long contexts or custom kernels will need to test their own workloads. The platform is also built on Arctic Platform, an open source backend the company said is fully visible and patchable by advanced teams.
Watch whether enterprises trade a bit of portability for that closed-loop convenience and whether Snowflake's governance controls and cost curve hold up once RL rollouts scale beyond demos.
