Recently, Xiaohongshu officially released the brand-new open-source large model dots3-note preview. As the first version of this series, the model has a total parameter count of 280B and an activated parameter count of 16B, and it possesses comprehensive multimodal perception and understanding capabilities for text, vision, and speech. In tasks such as reasoning, Agent, and multimodal processing, its performance is comparable to larger and more advanced models both domestically and internationally.
On the day of the model's release, Huawei quickly announced that the Atlas800A3 and Atlas900A3 SuperPoD nodes have been efficiently adapted to this model, and they provide complete deployment and operation support through the vLLM Ascend open-source inference engine.
This time, Ascend has carried out deep optimization throughout the entire pipeline, focusing on overcoming performance bottlenecks in multimodal processing and incremental decoding. In full multimodal end-to-end adaptation, the platform successfully connected the visual encoder, audio feature extraction, and the main reasoning of the large language model, ensuring the stability of native multimodal interaction. In terms of framework optimization, by deeply integrating communication and computation and reconstructing operators, repeated computations between multiple cards were completely eliminated, significantly improving the inference throughput of the MoE architecture.
In addition, to address the latency issues during the incremental inference phase, Ascend has natively adapted the MTP speculative decoding capability, which generates multiple candidate tokens in a single forward pass and completes verification and reuse, greatly reducing the number of decoding forward passes. These underlying optimizations not only ensure the efficient and stable operation of the model but also lay a solid computing foundation for the intelligent implementation of long-term and complex task scenarios.
