Arctic-TILT

Arctic-TILT is a sub-billion parameter, multimodal LLM delivering state-of-the-art Document Understanding (DU) performance comparable to models 1000x its size.

This is Arctic-TILT: a specialized, highly efficient model for enterprise Document Understanding, developed by Snowflake. It leverages a proprietary transformer architecture with unique modality fusion, seamlessly integrating text and visual data for unparalleled versatility. The key metric: Arctic-TILT achieves state-of-the-art results on seven diverse DU benchmarks, matching the accuracy of models 1,000 times larger. Critically, this sub-billion parameter model runs efficiently on a single, consumer-grade 24GB GPU (e.g., NVIDIA A10), drastically lowering operational costs. It handles long-context documents, processing visually rich files up to 400,000 tokens (approximately 500 pages), making it the definitive choice for cost-effective, large-scale document processing.

https://huggingface.co/Snowflake/snowflake-arctic-tilt-v1.3