Snowflake launches Arctic, an open model for enterprise SQL

Snowflake today introduced Arctic, an open language model designed for one of the most common tasks in business: turning questions and business needs into data queries, code and automations. The company is releasing it under the Apache 2.0 license, allowing it to be used, modified and redistributed for commercial purposes as well.

Arctic arrives as major data platforms seek to let customers interact with data warehouses using natural language. Generating a correct SQL query—the language used to query databases—is an especially valuable application: it can save analysts and developers time, but it requires the model to understand both the request and the specific structure of each company’s data.

480 billion total parameters, but 17 billion active

Arctic’s main technical feature is its mixture-of-experts architecture, known by the acronym MoE. Instead of activating the entire neural network for each word or piece of text, the model selects a small portion of its specialized components to handle each operation.

Arctic has 480 billion parameters in total, but activates roughly 17 billion for each token, the smallest unit of text the system processes. Snowflake built the model with 128 experts and activates two of them per token. The goal is to combine the capacity of a very large model with lower inference costs—the cost of using the model after training—than a dense network of equivalent size.

The distinction matters, although it does not eliminate the infrastructure requirements. Having only a fraction of the parameters participate in each response reduces the amount of computation needed, but hosting a 480-billion-parameter model still requires substantial memory and specialized hardware. For many companies, the practical option will remain using it through managed services or adapting smaller versions.

A model for queries, code and instructions

Snowflake has optimized Arctic for SQL generation, programming and instruction following, three capabilities that come together in internal data assistants. An employee might ask, for example, for sales in a region over a given period; the system would need to translate that request into SQL, use the actual table and column names, run the query with the appropriate permissions and explain the result.

The hardest part is not simply producing a syntactically sound SQL statement. In a real organization, the model must handle extensive schemas, business definitions that vary across departments and access rules for sensitive information. An apparently reasonable query can return the wrong figure if it confuses a metric or combines incompatible tables.

According to Snowflake, Arctic delivers competitive results on SQL, coding and instruction-following tests compared with other open models aimed at enterprise use. That is a meaningful claim, but benchmarks do not replace testing with each customer’s data and workflows. Public evaluations typically measure well-defined problems; a corporate database introduces ambiguity, incomplete data and permissions that a test cannot fully reproduce.

The license offers an alternative to closed models

The Apache 2.0 license puts Arctic in a different position from commercial models that can only be accessed through an API. A company can download the model’s weights, run it on its own infrastructure and fine-tune it for specific tasks. It can also review how it behaves within its own technology environment, which is important for industries with strict data-residency or auditing requirements.

However, opening up the model does not by itself solve data-governance problems. Companies will need to decide what information is sent to the system, limit which queries it can run, validate its responses and maintain controls over personal, financial or confidential data. An assistant that generates SQL should have fewer permissions than an administrative account, even when its results appear correct.

Arctic joins a growing race to develop open mixture-of-experts models. Mistral popularized the approach with Mixtral, and Databricks introduced DBRX in March with an MoE architecture and an Apache 2.0 license. Snowflake is now seeking to turn its position as a data platform into an advantage in the AI layer: not only storing and querying information, but also helping write the queries and applications that use it.

The next step will be to see how Snowflake integrates Arctic into its products and development tools, and whether the model maintains its performance outside the published benchmarks. For data teams, its appeal will not depend on the number of parameters, but on whether it reduces manual work without creating new avenues for errors, data leaks or queries that are difficult to audit.