Hugging Face download statistics

AI2 Launches Molmo: Open-Source AI Models in 1B, 7B, 72B Sizes, Using 1000x Less Data

Sep 25, 2024, 03:42 PM

The Allen Institute for AI (AI2) has launched Molmo, a family of open-source multimodal AI models. These models, available in 1B, 7B, and 72B-parameter sizes, are designed to outperform proprietary systems such as GPT-4V, Claude 3.5 Sonnet, and Gemini 1.5 Pro. Molmo models excel in vision and language tasks, leveraging a novel dataset called PixMo, which includes high-quality image-caption pairs and multimodal instruction data. The models are capable of rich interactions in both physical and virtual worlds, using 1000x less data compared to their closed-source counterparts. AI2's Molmo aims to democratize access to advanced AI capabilities by providing open weights and allowing researchers and developers to build upon them. Molmo also shows impressive performance on RealworldQA and OOD robotics perception tasks.