Snowflake's AI research: A look inside our generative AI innovations
summary
Snowflake open-sources Arctic under Apache 2 and shares concrete recipes: training on 3.5 trillion tokens to a mixture-of-experts architecture with 128 experts with top-2 routing, yielding ~17 billion active parameters optimized for SQL, coding, and instruction following. Tooling and kernels use DeepSpeed, Hugging Face, VM, TRT LLM, replicate and a Streamlit demo; custom kernels and activation quantization reduce compute and make serving across model gardens and Snowflake Cortex practical and functional.
AI-written summary · about
chapters (8)
- 0:00Snowflake AI research intro
- 3:31Arctic's enterprise focus
- 12:30Arctic pre-training data
- 36:11Arctic model architecture
- 52:07Fine-tuning Arctic
- 57:16Serving Arctic
- 1:03:50Arctic resources recap
- 1:09:22Q&A on data & performance
related talks
04mMaximize enterprise AI with seamless interoperability
49mLLM inference and fine-tuning on secure enterprise data
22mTraining Arctic
36mImplementing Natural Language Analytics with Snowflake Cortex
15mFrom Air Quality to Aircraft and Automobiles, Data Is Everywhere
29mOpen Source Sovereign AI: Building Enterprise GenAI Platforms for Control & Comp...
27mUsing GenAI with Audio Data for Driving Better Customer Experience
28mAutomating Database Operations with Cortex Code