Cloud data warehouse heavyweight Snowflake Inc. is looking to make its presence felt in the generative artificial intelligence arena with the debut of its own, open-source large language model, designed to compete with the likes of OpenAI’s GPT-4, Google LLC’s Gemini, Meta Platforms Inc.’s Llama 2 and Mistral AI’s Mixtral.

The company claims Snowflake Arctic is the most open, enterprise-grade LLM available today. What’s more, the company says, it’s also among the most powerful, aided by an up-and-coming “mixture of experts” architecture that optimizes it to perform a wide range of enterprise use cases.

Along with the model itself and some impressive benchmark results, Snowflake is also making Arctic’s weights available under an Apache 2.0 license, as well as details of how it was trained, setting what it says is a new standard for AI openness.

Snowflake’s new chief executive officer, Sridhar Ramaswamy (pictured), who replaced Frank Slootman in February, hailed the release of the Arctic LLM as a “watershed moment” for the company, which until now has been more invested in feeding data to LLMs rather than creating its own.

“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,” Ramaswamy said, adding that the aim is not so much to position Snowflake as an AI company as to enable its customers to use AI more more efficiently and reliably.

“We are making a foundational investment,” he told reporters in a briefing. “It’s our first big step in generative AI with lots more to come,” including at its Snowflake Summit event in early June in San Francisco.

Snowflake’s AI foundation

With its popular cloud data warehouse, Snowflake is already sitting at the heart of many enterprise’s generative AI initiatives, as the data it pools from various sources is widely used to train and fine-tune a variety of AI applications. In a report last month, Snowflake said an analysis of more than 9,400 customer accounts showed that AI use cases are increasingly prevalent among its user base, with the usage of the Python programming language, tags, unstructured data and purpose-built AI development tools all showing double- or triple-digit growth.

But Snowflake wants to do more than just feed AI with the data it needs to do its job. It said it wants to give customers more flexibility and choice around which LLMs they work with.

Dave Vellante, chief analyst at theCUBE Research, SiliconANGLE Media’s research unit, said Snowflake is in the midst of a transition catalyzed by the AI awakening and customer demand for open data platforms. Although it’s still a data company at heart, he believes the launch of Arctic signals it has a mandate to deliver end-to-end AI-enabled solutions.

“Does the world need another LLM?” Vellante asked. “Probably not but Snowflake needs one to demonstrate AI leadership and understand how to optimize its data platform for LLMs. At the same time, in my view, Snowflake must demonstrate proficiency with a diversity of LLMs and offer customers optionality.”

Together with its Apache 2.0 license that permits ungated personal, research and commercial use, Snowflake is also offering an assortment of code templates and flexible training and inference options for customers, so they can quickly deploy the LLM on their preferred frameworks. Options include Nvidia’s Corp.’s NIM with Nvidia TensorRT-LLM, vLLM and Hugging Face.

Additionally, Arctic is being made available for serverless inference within Snowflake Cortex, which is a fully-managed service for AI and machine learning that’s directly integrated with its flagship Data Cloud platform. “An open source model increases trust and transparency,” said Baris Gultekin, head of AI at Snowflake.

One interesting upshot of all the open-source or near-open-source models is that it’s now clear that generative AI won’t be a winner-take-all market such as search, Ramaswamy said. “AI is rapidly decentralizing,” he said.

Superior token efficiency

The LLM itself is one of the most powerful such offerings on the market, Snowflake insisted. Thanks to its differentiated MoE architecture, it significantly improves both training systems and model performance, and the company offers the benchmark results to back up those claims.

For instance, tests show that it activates 17 out of 480 billion parameters at a time to achieve “unprecedented token efficiency” during inference and training. That’s approximately half the amount of parameters used by Databricks Inc.’s open-source DBRX, and 80% fewer than the parameters used by xAI Corp.’s Grok-1. Snowflake says Arctic LLM also outperforms models such as Llama 2 70B, Mixtral-8x7B and others in coding benchmarks such as HumanEval+ and MBPP+, and SQL-generation tasks such as Spider and Bird-SQL, while demonstrating strong performance in terms of general language understanding.

What’s more, Arctic’s MoE architecture made it substantially easier to train, with Snowflake claiming that it was built from scratch in less than three months and at roughly one-eighth of the training costs of similar LLMs. By setting a new baseline for LLM training and costs, it will enable enterprises to build dramatically lower-cost generative AI models at scale, the company said.

Snowflake’s new CEO appears to be determined to ensure the company catches up with its rivals in the generative AI industry, having fallen behind its rival Databricks when that company launched its first LLM, called Dolly, just over a year ago, said Constellation Research Inc. analyst Doug Henschen.

Databricks further advanced its generative AI chops with the launch of the open-source DBRX in March, and that model appears to be getting a lot of traction, the analyst said. “Snowflake is clearly responding to the competitive threat, given the press release’s comparisons between the new Arctic LLM and DBRX,” Henschen stated.

While Snowflake claims some impressive benchmark performances, the analyst said he’s more interested to learn about the intended use cases. He also cautioned that Snowflake’s Cortex platform for AI, machine learning and generative AI development is still in private preview. “I’d like to see independent tests, but the breadth of customer adoption will be the ultimate gauge of Arctic’s success,” the analyst added. “As a customer, I would want to look beyond the performance claims and know more about vendor indemnification and risks when using LLMs in conjunction with RAG techniques.”

Vellante agreed that Snowflake is most definitely trying to one-up Databricks’ DBRX model, positioning it as a more open alternative with less restrictions on its use. “What matters to customers is gaining easy access to cost effective options that can be customized at scale,” he explained. “So all these LLMs, while sometimes confusing to buyers, means well-funded companies are investing like crazy in building products that solve problems. That is a real positive for enterprises.”

Andy Thurai, another analyst at Constellation Research, agreed on the positives, saying Arctic is a lot more open than other open-source models, such as the Llama series and Databricks’ DBRX. “It’s licensed under Apache 2.0, which permits ungated personal, research and personal use,” he explained. “This is different to other open-source LLMs, which only allow very limited use cases, with heavy restrictions on commercial use.”

The less restrictive nature of Arctic gives Snowflake a chance to catch up with its rivals, and the company can also leverage its strong data foundations, Thurai added. “If Snowflake can convince users to keep their data in its data lake to train their custom models, it can compete very well.”

Snowflake said Arctic has already attracted keen interest from a number of companies building generative AI applications, including the generative AI-based search engine startup Perplexity AI.

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

The Arctic LLM is a part of Snowflake’s wider family of Arctic AI models, which are also available under an Apache 2.0 license via platforms such as Hugging Face. The Arctic family includes the newly announced Arctic Embed, which is a text embedding model that’s optimized for data retrieval performance. Using Arctic Embed, companies will be able to set up retrieval augmented generation or RAG workflows to enhance their LLMs with their own proprietary datasets, and enable semantic search across those datasets.

In line with its open ethos, Snowflake reiterated that it continues to offer customers access to one of the largest selections of third-party LLMs, including the latest open-source offerings from Reka AI Inc. and Mistral AI. It also provides direct access to the compute infrastructure required to train and run LLMs, for example through its partnership with Nvidia.

With reporting from Robert Hof

Photo: SiliconANGLE; image: Snowflake