Qwen releases Qwen3.8-27B open-weights multimodal model

Qwen3.8-27B, a dense natively multimodal model (images + video), shipped with FP8 weights on HuggingFace, followed by community GGUF quants from unsloth, bartowski, and huihui-ai. llama.cpp PR #27342 adds dflash2 decoding, reported up to 140.6 tok/s on an RTX 6000.