Moondream: Moondream 3 Preview

Moondream 3 Preview, a 9B parameter vision-language model with a 2B active mixture-of-experts architecture, demonstrates exceptional speed and cost-efficiency. It consistently ranks among the fastest models and offers highly competitive pricing across all evaluated benchmarks. However, its performance in specific cognitive tasks shows significant room for improvement. In Instruction Following, the model achieved 19.0% accuracy, placing it in the 18th percentile, indicating challenges with complex multi-layered instructions. Performance in Coding and Reasoning benchmarks was notably low, with 0.0% accuracy in both categories, suggesting that these areas are significant weaknesses. While the model excels in operational metrics like speed and cost, its current capabilities in complex problem-solving and knowledge-based tasks are limited. This model appears well-suited for applications where rapid processing and low cost are paramount, and where the complexity of visual understanding or reasoning is less demanding.