Snowflake Cortex Fine-Tuning
SnowflakeSnowflake Cortex Fine-Tuning is a fully managed function inside Snowflake Cortex AI that applies parameter-efficient fine-tuning (PEFT) to a fixed list of base models: Llama 3 (8B/70B), Llama 3.1 (8B/70B), Mistral 7B, and Mixtral 8x7B. Training data comes from tables in the customer's account, and both training and inference run on Snowflake compute rather than a third-party platform, billed per token in credits.
Verified 2026-08-09 via the Snowflake Cortex Fine-tuning documentation.
Openness
1 high confidence- license
- Proprietary
- service
- fully-managed-proprietary(in-Snowflake)
- source
- closed
- method
- PEFT(LoRA-class, not specified)
- base-models
- fixed allowlist
- access
- Snowflake-account+credits
- data
- user-data stays in account but engine closed
A fully managed, proprietary fine-tuning feature inside Snowflake, with no source published and no way to self-host.
- https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-finetuning recorded 2026-08-09
Cortex Fine-tuning is a fully managed service that runs entirely inside Snowflake via the FINETUNE SQL function or the Snowsight AI Studio wizard; the page names no published repository, no self-host option, and no license terms for the underlying engine, and exposes only a fixed allowlist of six base models (llama3-8b, llama3-70b, llama3.1-8b, llama3.1-70b, mistral-7b, mixtral-8x7b) through the managed API.
Adoption
not assessedNo usage or customer-count figure is published for the fine-tuning feature specifically. Snowflake has a large data-platform install base, but nothing about it can honestly be read as a signal for how much this feature is used, so no level is assigned.
- https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-finetuning recorded 2026-08-09
Documents a CORTEX_FINE_TUNING_USAGE_HISTORY account_usage view for tracking a customer's own credit and token consumption, but discloses no aggregate customer count, deployment count, or other usage figure for the fine-tuning feature itself.
Capability
2 medium confidenceNarrow on every axis this category compares: a single parameter-efficient method of the LoRA class, over a fixed allowlist of six base models (Llama 3 and 3.1, Mistral and Mixtral), with no full fine-tuning, no RLHF, DPO or GRPO, and no user choice of model or training scale. Integrated convenience for Snowflake users rather than a frontier trainer.
- https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-finetuning recorded 2026-08-09
States that Cortex Fine-tuning lets users leverage parameter-efficient fine-tuning (PEFT) to create customized adaptors, over a fixed set of six base models (llama3-8b, llama3-70b, llama3.1-8b, llama3.1-70b, mistral-7b, mixtral-8x7b); no full fine-tuning, RLHF, DPO, or GRPO option is mentioned anywhere on the page.
Verified 2026-08-09