Xiaohongshu has released the first version of its open-source large model series dots3-note, featuring 280B total parameters and 16B activated parameters, with strong multimodal perception and understanding capabilities across text vision and speech, and performance comparable to larger international models in reasoning Agent and multimodal processing tasks. On the release day Huawei announced it has adapted its Atlas800A3 and Atlas900A3 SuperPoD nodes for this model, providing full deployment and operation support via the vLLM Ascend open-source inference engine. Ascend optimized the entire pipeline (multimodal processing incremental decoding framework) to eliminate performance bottlenecks and boost inference throughput in MoE architecture. It also natively integrated MTP speculative decoding to cut decoding forward passes, ensuring efficient stable model operation and supporting complex long-term task execution.