RedNote open-sources dots3-note, Huawei Ascend adapts full-modal model

RedNote has officially released the dots3-note preview open-source large model, the first version of the dots3 series. With 280B total parameters and 16B active parameters, it supports full-modal understanding of text, vision, and audio. Huawei's Ascend team completed 0-day adaptation on the release day, offering full deployment via the vLLM Ascend inference engine.

With FlashComm communication optimization and FUSED_MC2 compute-communication fusion, Ascend boosts multi-card inference efficiency. It also natively supports MTP speculative decoding, cutting per-token latency and supporting high-concurrency scenarios. The model reportedly rivals larger models on multiple tasks.