Meet Snowflake Arctic, our new LLM!

A truly open large language model that pushes the frontiers of cost-effective training and openness.

Posted in LLMs,
Meet Snowflake Arctic, our new LLM!

👋 Dear beloved Streamlit Community!

We have some pretty big news for you today. Here’s the tea. 🍵

Today, Snowflake is launching a new foundation language model, Snowflake Arctic, which is super efficient and intelligent. Arctic was in part inspired by Streamlit. It was designed for Streamlit too! And we think the Streamlit community has an important (and fun!) role to play in Arctic’s future! Let’s unpack. 

What is a foundation language model?

Like OpenAI’s GPT-4, Anthropic’s Claude, or Mistral’s Large model, foundation language models form the brains behind generative AI. With billions of parameters encoding vast amounts of “world knowledge,” these neural nets let you ask questions, have conversations, or reason about problems. 

Efficient intelligence.

Arctic is on-par or better at enterprise grade tasks than industry leading open source foundational models, at a fraction of the compute cost. It is also vastly more intelligent than other open source models trained with a similar compute budget. Snowflake’s AI Research team developed some truly novel methods to achieve unprecedented training and inference efficiency. I absolutely recommend reading the full blog from the stellar Snowflake AI Research team where they go into all the important details about how Snowflake Arctic was built, how it compares to other models, and why it’s so good at what it does.

Truly open.

Inspired by Streamlit, Snowflake Arctic is truly open source. When we launched Streamlit, we decided to use the ultra-permissive Apache-2.0 license in order to make our tools as widely available as possible and create a rich community of users. Today, the Streamlit community comprises 275K monthly active developers (you!), tons of community-created documentation, and lots of excitement and energy. Just two weeks ago, over 700 of you spent a full hour live with us to discuss Streamlit’s roadmap.

This is why we decided to open-source Snowflake Arctic too – giving you access to model weights, code, and training recipes under the exact same Apache-2.0 license as Streamlit itself. This gives you all the transparency needed to fully evaluate and trust Arctic, as well as to create your own cool derivative works. 

Designed to help you build with Streamlit.

Arctic is just amazing at writing Streamlit code. In fact, Snowflake Arctic scores a 0.643 average on code generation metrics HumanEval+ & MBPP+, making it a leading code-generating model in its class. This opens up a huge set of possibilities, like auto-generating Streamlit apps from prompts, or auto-inlining snippets of Streamlit in a chat conversation

Here’s an example of just that:


Which brings us to you!

This is just the beginning. You, the Streamlit community, have many more ideas and creative applications than we do! What can you do with Arctic that we haven’t even thought of? We can’t wait to find out! 

How you can get started today:

  • 💬 Chat with Arctic! Try a live demo now on Streamlit Community Cloud or on Hugging Face Streamlit Spaces, with an API powered by our friends at Replicate.
  • 🚀 Remix our starter code on GitHub to build your own magical apps with Arctic!
  • 💻 Register for our Arctic-themed Streamlit Community Hackathon, and get mentorship plus exclusive platform access to help you build your own Arctic powered magical apps. 
  • 👉 Arctic is coming to your model garden or catalog of choice soon, including AWS, NVIDIA AI Catalog, Lamini, Perplexity, Replicate and Together AI!

Don’t forget to share your icy cool Artic apps on the Streamlit Community Forum or social media with the hashtag #SnowflakeArctic

Can’t wait to see what you all build!

❤️ Adrien

Share this post


Continue the conversation in our forums →

Also in LLMs...

View even more →