Qwen releases FP8 27B vision-language model with 262K native context and agent benchmark gains · cho.sh