Snowflake

Data infrastructure with an AI research contribution

Snowflake combines storage, compute and cloud services for analytics, data engineering, applications and AI workloads. This places it both in the systems that organize enterprise data and in the tooling used to build models and applications around that data. [1]

Arctic weights and inference implementations

The April 2024 Arctic release provided base and instruction-tuned checkpoints. The model card identifies a hybrid mixture-of-experts architecture with 480 billion total and 17 billion active parameters and an Apache 2.0 license. This is a weights release, distinct from access to a particular hosted endpoint. [3]

Snowflake credited the NVIDIA TensorRT-LLM and vLLM teams with preliminary inference implementations. That is a specific engineering contribution. The release’s claimed efficiency and benchmark advantages remain Snowflake’s measurements, rather than universal results for every deployment. [2]