Super cool models! Love the mixture of sliding window and global attention to cut down on KV cache sizes! And 4B, 12B and 27B are vision + text by default! Super cool!