Snowflake unveiled its latest artificial intelligence (AI) innovation: the Arctic large language model (LLM). With its mixture-of-experts (MoE) architecture, the model is optimized for complex enterprise workloads to offer efficiency at scale. In a move toward openness, the company is releasing Arctic’s weights under an Apache 2.0 license and details of the research leading to how it was trained.

“We think that Snowflake is the most important enterprise AI company on the planet, because we are the data foundation,” Snowflake CEO Sridhar Ramaswamy said during the press briefing. “Snowflake Arctic ... is a huge milestone for AI innovation, and a big step forward for enterprise-grade, open large language models.”

Baris Gultekin, head of AI at Snowflake, echoed that the launch of Arctic is a pivotal moment for the company. “Snowflake is setting a new baseline for how fast and efficient the state-of-the-art, open-source models can be trained, and ultimately enabling customers to create cost-efficient custom models at scale.”

He touts the Snowflake Arctic has three differentiators: First, it delivers top-tier enterprise intelligence; second, the new LLM tops the benchmarks for complex enterprise workloads such as SQL generation, code generation and instruction following and achieves this through efficiency with a MoE architecture; Lastly, Arctic LLM is an open model.

The Snowflake Arctic LLM is a part of the Arctic model family that also includes practical text-embedding models for retrieval use cases.

Snowflake highlights the openness of Arctic LLM

Snowflake claims the Arctic LLM is an open model with an Apache 2.0 license that permits ungated personal, research and commercial use. The permissive license is commercially available, Gultekin said.

In addition, the data lake and analysis company also provides code templates, flexible inference and training options so users can deploy and customize Arctic LLM using their preferred frameworks, including Nvidia NIM with Nvidia TensorRT-LLM, vLLM and Hugging Face.

“Not only are we making the model available, but also making all of the code available for fine-tuning in a very efficient way,” Gultekin said.

“Customers want reassurances about their models, they want to be able to train their models with their own data. So we believe by putting an open source model out there, it increases trust, it increases transparency,” he added.

Building enterprise AI with resource-efficiency

Another aspect of Arctic's development Snowflake highlighted is its cost-efficiency.

“Snowflakes research team has taken less than three months and spent roughly one eight the training costs of similar models while building the Arctic. And we believe this translates directly to our customers as they build their custom models,” Gultekin said.

Snowflake claims the Arctic activates 17 out of 480 billion parameters at a time to achieve industry-leading quality with token efficiency, which is about 50% fewer parameters than its rival Databricks LLM DBRX, and 80% less than X.ai’s LLM Grok-1 during inference or training.

“Arctic is on par or better than [Meta’s] Llama 380 or the Code lLama on all metrics while using less than two times to training compute budget,” Gultekin said. “But similarly, despite using 17x less compute budget, Arctic is on par with Llama3 70B in language understanding and reasoning while surpassing in Enterprise Metrics, for example, SQL generation, coding, and instruction following.”

“This high training efficiency of Arctic also means that users and snowflake customers can train custom models in a much more affordable way,” he added.