“The era of enterprise artificial intelligence is here,” states in the preamble Sridhar Ramaswamy, appointed at the end of February to the position of CEO of Snowflake. And it is clear that the data cloud specialist is not immune to the trend which consists of developing its own large language model (LLM). The company thus unveiled Arctic, an open source and efficient LLM.
Based on a Transformer Dense-MoE (mix of experts) architecture, Arctic is optimized for complex enterprise workloads as well as SQL code generation and statement tracking. This is a combination of a 10B dense Transformer model with a residual MLP of 128×3.66B MoE, resulting in 480 billion total parameters and 17 billion chosen active parameters. It was designed with a pop-up of 4000 tokens.
Advertisement
A publication under Apache 2.0 license
The generative AI model is part of a family of eponymously named models built by Snowflake. As a demonstration of its commitment to the open source community, Snowflake has released the weights and Arctic code under an Apache 2.0 license (which allows unrestricted personal, research, and commercial use) along with the research details that led to his training.
Snowflake wants to deploy its LLM on a large scale
“This is a watershed moment for Snowflake, our innovative AI research team at the forefront of AI,” said Sridhar Ramaswamy, CEO of Snowflake. The latter is also very ambitious and does not hesitate to maintain that thousands of customers will soon exploit the potential of Arctic. “As the database for more than 9,400 companies and organizations worldwide, Snowflake enables all users to leverage their data with leading open LLMs, while providing flexibility and choice of models with whom they work. Its clients include Adobe, Allianz, AXA, Capital One, KFC, Logitech, Mastercard, Micron, Rakuten, Strava, and Western Union.
Going further, Snowflake also provides code templates, as well as flexible inference and training options so users can quickly start deploying and customizing Arctic using their preferred frameworks. These include Nvidia NIM with TensorRT-LLM, vLLM and Hugging Face. For immediate use, Arctic is available for serverless inference in Snowflake Cortex, its fully managed service that delivers ML and AI solutions in the Data Cloud. Arctic will also be available on Amazon Web Services, alongside other models and catalogs, which will include Hugging Face, Lamini, Microsoft Azure, Nvidia API Catalog, Perplexity and Together AI to name a few.
An LLM developed in less than three months at low cost
Snowflake's AI research team says it took less than three months and spent about an eighth of the cost of training similar models when developing Arctic – “less than two million dollars according to Sridhar Ramaswamy”. Trained using Amazon EC2 P5 instances, Snowflake hopes to set a new benchmark for how quickly open models can be trained.
Advertisement
The performance/cost ratio is in fact relevant for both the training part and the inference part and the models developed by Databricks and Meta seem far from the mark. Compared to Llama 3 70B – recently released – Snowflake reports that the multiplier is close to 16 in terms of computing power used to train the model compared to its own.
Arctic thus provides quality results, activating only 17 of the 480 billion parameters at a time. In terms of efficiency, Snowflake claims that Arctic activates approximately 50% fewer parameters than DBRX, and 75% fewer than Llama 3 70B during inference or training.
Performance that matches Llama 3
On major benchmarks, Snowflake LLM outperforms leading open models. Thus, Arctic obtains a score of 79%, tied with Llama 3 70B and Mixtral 8x22B on SQL generation (Spider test) and close behind the Meta and Mistral AI models in terms of coding (HumanEval+, MBPP+). Evaluations which highlight the lower performance of the open source DBRX model – published by Databricks recently – already lagging behind its competitors.
On the academic side, Arctic offers interesting performances overall, particularly in mathematics (on the GSM8K test), but is slightly behind in terms of general language understanding (MMLU) – finding itself behind DBRX, Llama 3 70B , Mixtral 8x22B and Mixtral 8x7B. In addition to LLM Arctic, the Snowflake Arctic template family also includes Arctic embed, a family of text embedding templates, recently announced and also released under the Apache 2.0 license.
The open source ecosystem is consolidating
In total, the family, which contains five models, is available on Hugging Face for immediate use and will soon be available as part of the Snowflake Cortex integration feature (in private preview at this time). “These integration models are optimized to deliver industry-leading search performance at approximately one-third the size of comparable models,” Snowflake advances.
Remember that at the same time, the data cloud specialist has made several strategic mergers in the artificial intelligence sector. It has notably invested in the French start-up Mistral AI and offers the latter's models on its platform. Similarly, Snowflake works with American start-up Perplexity and even uses its search service internally.
For Aravind Srinivas, co-founder and CEO of Perplexity, the announcement made today by the data cloud specialist is a step forward towards the consolidation of the open source ecosystem: “The continued advancement of – and healthy competition between – open source AI models is essential not only to the success of Perplexity, but also to the future of democratizing generative AI for all.” And to add: “We look forward to experimenting with Snowflake Arctic to adapt it to our product.”
Do you want to stay up to date on the latest news in the artificial intelligence sector? Register for free to the IA Insider newsletter.
Selected for you