After open-sourcing the Arctic family of text embedding models, Snowflake is now adding another LLM to the list for enterprise use cases. Snowflake Arctic sets a new standard for openness and enterprise-grade performance. 

Designed with a unique Mixture-of-Experts (MoE) architecture, Arctic provides top-tier optimisation for complex enterprise workloads, surpassing several industry benchmarks in SQL code generation, instruction following, and more. 

Arctic’s unique MoE design enhances both training systems and model performance with a carefully crafted data composition tailored to enterprise needs. With a breakthrough in efficiency, Arctic activates only 17 out of 480 billion parameters at a time, achieving industry-leading quality with unprecedented token efficiency.

“Despite using 17x less compute budget, Arctic is on par with Llama3 70B in language understanding and reasoning while surpassing in Enterprise Metrics,” said Baris Gultekin,  Snowflake’s head of AI.

Compared to other models, Arctic activates approximately 50% fewer parameters than DBRX, and 80% fewer than Grok-1 during inference or training. Moreover, it outperforms leading open models such as DBRX, Llama 2 70B, Mixtral-8x7B, and more in coding (HumanEval+, MBPP+) and SQL generation (Spider and Bird-SQL), while also providing superior performance in general language understanding (MMLU).

“This is a watershed moment for Snowflake, with our AI research team innovating at the forefront of AI,” said Sridhar Ramaswamy, CEO, Snowflake. “By delivering industry-leading intelligence and efficiency in a truly open way to the AI community, we are furthering the frontiers of what open source AI can do. Our research with Arctic will significantly enhance our capability to deliver reliable, efficient AI to our customers,” he said. 

The best open model?

The best part is that Snowflake is releasing Arctic’s weights under an Apache 2.0 licence, along with details of the research behind its training, establishing a new level of openness for enterprise AI technology. “With the Apache 2 licensed Snowflake Arctic embed family of models, organisations now have one more open alternative to black-box API providers such as Cohere, OpenAI, or Google,” says Snowflake.

“The continued advancement and healthy competition between open source AI models is pivotal not only to the success of Perplexity, but the future of democratising generative AI for all,” said Aravind Srinivas, co-founder and CEO, Perplexity. “We look forward to experimenting with Snowflake Arctic to customise it for our product, ultimately generating even greater value for our end users.”

As part of the Snowflake Arctic model family, Arctic is the most open LLM available, allowing ungated personal, research, and commercial use with its Apache 2.0 licence. Snowflake goes further by providing code templates, along with flexible inference and training options, enabling users to deploy and customise Arctic quickly using their preferred frameworks, including NVIDIA NIM with NVIDIA TensorRT-LLM, vLLM, and Hugging Face.

Yoav Shoham, co-founder and co-CEO, AI21 Labs, said, “We are excited to see Snowflake help enterprises harness the power of open source models, as we did with our recent release of Jamba — the first production-grade Mamba-based Transformer-SSM model.”

For immediate use, Arctic is available for serverless inference in Snowflake Cortex, Snowflake’s fully managed service offering machine learning and AI solutions in the Data Cloud, alongside other model gardens and catalogues such as Hugging Face, Lamini, Microsoft Azure, NVIDIA API catalogue, Perplexity, Together, and more.

“We’re pleased to increase enterprise customer choice in the rapidly evolving AI landscape by bringing the robust capabilities of Snowflake’s new LLM model Arctic to the Microsoft Azure AI model catalogue,” said Eric Boyd, corporate vice president, Azure AI Platform, Microsoft.

Everyone loves the winter

Snowflake’s AI research team, comprising industry-leading researchers and system engineers, developed Arctic in less than three months, spending roughly one-eighth of the training cost of similar models. Snowflake has set a new benchmark for the speed at which state-of-the-art open, enterprise-grade models can be trained, enabling users to create cost-efficient custom models at scale.

Clement Delangue, CEO and co-founder of Hugging Face said, “We’re excited to see Snowflake contributing significantly with this release not only of the model with an Apache 2.0 licence but also with details on how it was trained. It gives the necessary transparency and control for enterprises to build AI and for the field as a whole to break new grounds.”

Snowflake Ventures has also recently invested in LandingAI, Mistral AI, Reka, and others, reinforcing its commitment to helping customers derive value from their enterprise data with LLMs and AI. 

“Snowflake and Reka are committed to getting AI into the hands of every user, regardless of their technical expertise, to drive business outcomes faster,” said Dani Yogatama, co-founder and CEO, Reka. “With the launch of Snowflake Arctic, Snowflake is furthering this vision by putting world-class truly-open large language models at users’ fingertips.”

Additionally, Snowflake has expanded its partnership with NVIDIA to further AI innovation, combining the full-stack NVIDIA accelerated platform with Snowflake’s Data Cloud to provide a secure and powerful infrastructure and compute capabilities for unlocking AI productivity.