Snowflake
Snowflake's AI layer is Cortex - LLM functions, embeddings, and search that run as SQL functions directly on data already in Snowflake. No separate infrastructure, no data movement, no Python environment required for the common cases. Snowpark ML sits alongside Cortex for teams that need full model training and a model registry.
When to choose Snowflake: data already lives in Snowflake, the team building AI features is SQL-centric (analysts, BI engineers) rather than ML engineers, and the use case is summarization, sentiment, translation, or RAG search over governed data - not custom model training or complex multi-agent orchestration.
Architecture
Cortex AI provides managed model access; Snowpark ML, the Model Registry, and Notebooks cover custom model work. Both layers feed Search & RAG, where Cortex Search and the native VECTOR type turn stored data into retrieval-ready content.
Built-in AI Capabilities
Ratings out of 5 (5 = most mature). Reflects capability maturity as of mid-2026, not overall product quality.
| Category | Capability | Rating | Notes |
|---|---|---|---|
| Data Science & ML | Notebooks | 4/5 | Snowflake Notebooks, improving rapidly |
| Data Science & ML | Feature Store | 4/5 | Recently GA |
| Data Science & ML | Training | 4/5 | Snowpark ML, container services |
| Data Science & ML | Registry | 4/5 | Model Registry available |
| Data Science & ML | MLOps | 3/5 | Less mature than Databricks - no equivalent to MLflow's full lifecycle tooling |
| Generative AI | LLM Access | 4/5 | Cortex LLM functions, SQL-accessible; fewer model options than hyperscaler platforms |
| Generative AI | Embedding Models | 5/5 | Cortex embed functions - arctic-embed and multiple other models |
| Generative AI | Vector Search | 4/5 | Native VECTOR type, Cortex Search |
| Generative AI | RAG Support | 4/5 | Cortex Search built for RAG, simple SQL interface |
| Generative AI | Fine-Tuning | 3/5 | Limited fine-tuning options |
| Agentic AI | Agent Framework | 3/5 | Via Snowpark plus external frameworks; no native agent service |
| Agentic AI | Tool Calling | 4/5 | Cortex Analyst, Snowflake functions |
| Agentic AI | Orchestration | 3/5 | Manual implementation; no native agent orchestration |
| Agentic AI | Guardrails | 4/5 | Cortex Guard for content safety |
The pattern across all three categories: Snowflake is strongest where AI stays close to SQL (embeddings, vector search, tool calling) and weakest where a use case needs full lifecycle ownership - MLOps, agent orchestration, fine-tuning.
Available Models
| Type | Models | Notes |
|---|---|---|
| LLMs | Llama 3.1 (8B, 70B, 405B), Mistral Large, Mixtral 8x7B, Gemma, Snowflake Arctic | Accessed via CORTEX.COMPLETE() - no deployment step |
| Embeddings | snowflake-arctic-embed, e5-base-v2, multilingual-e5-large | Accessed via CORTEX.EMBED_TEXT_*() |
| Specialized functions | Sentiment, Summarize, Translate, Extract Answer | Purpose-built SQL functions, no model selection required |
Custom AI Enablement
The defining trait of Cortex: AI runs as a SQL function on data that never leaves Snowflake. No endpoint to provision, no client library to install, no infrastructure to manage - a query is the integration.
SELECT SNOWFLAKE.CORTEX.COMPLETE(
'llama3.1-70b',
'Summarize this customer feedback: ' || feedback_text
) AS summary
FROM customer_feedback;
Two other Cortex services extend this pattern beyond single-function calls:
- Cortex Analyst - natural language to SQL over a defined semantic model. Business users ask questions in plain English; Cortex Analyst generates and runs the SQL, grounded in the semantic model's metric and join definitions.
- Cortex Search - managed hybrid (vector and keyword) retrieval for RAG. Indexes a table or stage, handles embedding and ranking, and returns results ready to pass into a
COMPLETE()call - a full RAG pipeline built from two SQL-callable services.
Cost Model
Cortex functions bill on Snowflake credits, consumed per function call rather than against a fixed compute allocation. That is convenient at low volume - no warehouse to size specifically for AI workloads - but at high call volume, per-function cost can outpace the economics of a fixed-throughput inference deployment. Model expected call volume against CORTEX.COMPLETE() and similar functions before committing to Cortex as the primary path for a high-throughput production workload.
See official pricing before scoping: snowflake.com/en/data-cloud/pricing-options
Snowflake vs Databricks
Both platforms put AI next to the data, but they optimize for different teams.
| Capability | Snowflake | Databricks |
|---|---|---|
| Primary user | SQL analysts, BI engineers | ML engineers, data scientists |
| Barrier to entry | Low - SQL functions, no infrastructure to stand up | Higher - Python/Spark, MLOps setup |
| MLOps maturity | 3/5 - basic registry, no full lifecycle tooling | Mature - MLflow is the industry standard |
| Agent orchestration | 3/5 - no native agent service | Stronger, via Mosaic AI Agent Framework |
| Fine-tuning / open models | 3/5 - limited | Strong - full control over training and open-weight models |
| Best fit | Teams building AI features on data already in Snowflake, without ML engineering depth | Teams that own the full ML lifecycle and need agent orchestration or custom training |
The honest trade-off: Snowflake gets a SQL-centric team from zero to a working AI feature - summarization, sentiment, RAG search - faster than any other platform in this comparison. Databricks is the better choice once requirements move into full MLOps, native agent orchestration, or fine-tuning open-weight models. Cortex is not built for that depth yet.
Key Takeaways
- Cortex functions are SQL-callable -
COMPLETE(),SENTIMENT(),TRANSLATE(),SUMMARIZE()run directly inside a query. This is the platform's main differentiator: AI without a Python environment or infrastructure to provision. - Cortex Search and Cortex Analyst combine to form a full RAG and natural-language-to-SQL stack, both callable from SQL.
- MLOps, agent orchestration, and fine-tuning are the weakest areas (3/5 each) - budget for external tooling, or plan a handoff to Databricks, if the roadmap needs full model lifecycle ownership.
- Per-function credit billing is convenient at low volume and can get expensive at high call volume. Model the cost curve before committing Cortex to a high-throughput production path.
- Choose Snowflake's AI layer when the team is SQL-centric and the data already lives there. Choose Databricks when the team owns the full ML lifecycle.