Huawei Ascend Fully Adapts to Xiaohongshu's Dots3-Notepreview Large Model

On August 14, Xiaohongshu released dots3-notepreview, an open-source large model with a total of 280 billion parameters and 16 billion activated parameters, supporting full-modal perception and understanding across text, vision, and speech. On the same day, Huawei announced that its Ascend Atlas 800 A3 and Atlas 900 A3 SuperPoD hyper-nodes have completed full adaptation for this model and now offer complete deployment and inference capabilities based on the vLLM-Ascend open-source inference engine. The adaptation was finalized on the same day as the model’s release, featuring key optimizations including end-to-end full-modal integration, deep framework-level enhancements, and native support for MTP speculative inference—significantly boosting inference throughput and generation efficiency to meet the low-latency demands of high-concurrency online services.