If you want to know what a model actually does, skip the announcement and read the chat template. The Qwen3.8-27B-FP8 page proves the point: all it shows is the template, and that still says more than most launch posts — vision tokens for images and video, thinking on by default, and a reasoning_effort parameter with three levels: xhigh (the default), medium, and low.
That knob is what I’d wire up first. On self-hosted hardware, thinking tokens are latency and watts. A template-level setting that tells the model to keep its thinking brief and move directly to the conclusion means I can default to low for routine calls and only pay for xhigh when correctness matters.
One migration note: tool calls come out as XML function blocks, not JSON. If your tooling parses JSON tool calls, budget a shim before this touches anything real.
The story — Qwen has published Qwen3.8-27B-FP8 on Hugging Face. The visible page content is the model’s chat template, which accepts text, image, and video inputs, enables thinking by default with a reasoning_effort of xhigh, medium, or low, includes a preserve_thinking option for retaining reasoning content in the conversation history, and instructs the model to format tool calls as XML-tagged function blocks. (Source)