Dive Brief:
- Snowflake entered the generative AI model-building fray Wednesday with the release of Arctic, an open-source LLM optimized for SQL generation, coding assistance and natural language prompts.
- The data cloud company prioritized transparency and efficiency in the model. Snowflake engineers spent less than $2 million training Arctic in under three months, the company’s Head of AI Baris Gultekin, said during a Monday virtual briefing.
- Snowflake released the open-source model under a permissive Apache 2.0 license and published a research “cookbook” containing technical specifications and a recipe for fine-tuning Arctic on a single GPU. “Not only are we making the model available, but we’re also making the code available for fine-tuning,” Gultekin said.
Dive Insight:
AI aspirations are intimately tied to enterprise data capabilities. The technology is tethered to Snowflake’s strategy, too.
The data cloud company elevated in-house AI development with the February appointment of Sridhar Ramaswamy to CEO. The executive joined Snowflake as SVP of AI last May when the company acquired Neeva, an AI search engine startup he co-founded.
Ramaswamy was instrumental in Snowflake’s November launch of the Cortex AI managed service. Cortex is just days away from general availability, Ramaswamy said during the briefing Monday.
Arctic is one of a family of five incrementally sized models soon to be generally available via Cortex. The LLMs are currently deployed in the Hugging Face model marketplace. The company also brought its AI-powered SQL Copilot tool into public preview this month.
Snowflake’s model suite joins an increasingly crowded field of enterprise AI solutions.
Microsoft introduced the first of its Phi-3 small language models Tuesday. The economy-size Phi-3-mini is designed to run in compute-limited inference environments, including AI-optimized laptops and devices, the company said in a release.
Facebook parent company Meta delivered the Llama-3 cohort of open-source models Thursday. Unlike Arctic, Meta’s models did not include accompanying research, which the company said is forthcoming, and do not carry an Apache 2.0 license, a common open-source standard.
Despite Snowflake’s inroads into model building, Ramaswamy stressed the company’s keen focus on managing the enterprise data behind the AI boom.
“Snowflake is a data company — the majority of our resources still are devoted to how do we make the data core strong,” he said. “However, AI is an important layer that accelerates the value our customers are going to get from their data investments.”
