Tons of labs sloppily ship stuff nowadays, Laguna similarly had issues with their S-2.1 weights at launch and was later fixed seemingly in a similar post-training change, and almost always there is some issues in prompt templates at launch in various inference engines, even for "official" releases where the labs themselves contribute the patches. Google, Qwen (Alibaba), DeepSeek, Meta, and OpenAI, they're all guilty of doing sloppy releases in the past.