Snowflake Arctic: A New Open Source LLM
Snowflake, known for its efficient database, is rolling out a new LLM!
Created on April 24|Last edited on April 24
Comment
Snowflake has recently released Snowflake Arctic, a new language model (LLM) designed for enterprise use that emphasizes cost efficiency. This model is engineered to handle the rigorous demands of enterprise environments effectively, focusing on providing reliable performance while remaining accessible and easy to integrate.
Efficiency and Accuracy on Common Tasks
Snowflake Arctic distinguishes itself through its exceptional ability to handle complex enterprise tasks such as SQL generation, coding, and instruction following, which are crucial for applications like SQL data copilots and code assistance tools. Remarkably, it achieves these capabilities using a training budget significantly lower than traditional models, costing under $2 million in total. This represents a fraction of the usual expenses, which often stretch into the tens to hundreds of millions of dollars.
The model's architecture, a Dense-MoE Hybrid transformer, combines a 10B dense transformer with a residual 128×3.66B MoE MLP, resulting in a total of 480B parameters but only 17B active parameters during operations. This unique setup provides a balance between high capacity and resource efficiency, leading to lower training and operating costs.
Performance Metrics
Snowflake Arctic is rigorously tested across various benchmarks to ensure its effectiveness and efficiency:
Enterprise Intelligence Metrics: These include Coding (HumanEval+ and MBPP+), SQL Generation (Spider), and Instruction Following (IFEval). Arctic's performance in these metrics is exemplary, proving its aptitude for complex enterprise tasks.
Academic Benchmarks: It also excels in a broader range of metrics like Language Understanding and Reasoning, which include tests from ARC-Easy to WinoGrande, and Math (GSM8K), demonstrating strong general AI capabilities.
Despite its efficient use of computational resources, Arctic compares favorably with more resource-intensive models, demonstrating its potential to provide top-tier enterprise intelligence at a fraction of the usual cost.
Availability
Licensed under Apache 2.0, the model allows ungated access to model weights and code, encouraging collaboration and innovation within the AI community. Additionally, Snowflake has open-sourced various training optimizations and system enhancements, further contributing to the field.
The model is readily accessible through platforms such as Hugging Face, Amazon Web Services, Microsoft Azure, and NVIDIA API catalog, ensuring that developers and enterprises can easily integrate Arctic into their systems.
The Future
Looking ahead, Snowflake's vision for Arctic includes not only enhancing its capabilities but also expanding its accessibility and community engagement. The upcoming Data Cloud Summit and community hackathons are just a few of the platforms where Snowflake aims to engage with users and developers, driving further advancements in enterprise AI.
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.