Alibaba Releases Qwen3.8-27B-FP8 Model on Hugging Face
Original: Qwen 3.8 27B
Why This Matters
A 27B FP8 model with adjustable reasoning and multimodal support lowers the deployment barrier for capable open-weight AI.
Alibaba's Qwen team has published Qwen3.8-27B-FP8, a 27-billion-parameter language model in FP8 quantization format, on Hugging Face. The model supports configurable reasoning effort levels and multimodal inputs including images and video.
Alibaba's Qwen team has released Qwen3.8-27B-FP8 on Hugging Face, a 27-billion-parameter model stored in FP8 (8-bit floating point) precision to reduce memory and storage requirements compared to full-precision variants. The model's chat template reveals support for adjustable reasoning effort — configurable as 'xhigh', 'medium', or 'low' — allowing users to trade off compute against answer quality. At 'xhigh', the model is instructed to 'think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness'; at 'low', it is directed to keep reasoning brief. The template also exposes multimodal capabilities, handling both image and video inputs via vision tokens, though system messages are restricted to text only. Tool-calling is supported through a structured XML-based format requiring function calls to be wrapped in specific tags. The FP8 quantization format makes the model more accessible for deployment on consumer and enterprise hardware with limited GPU memory, while retaining most of the full-precision model's capability.