Huawei's Ascend adds 0-day support for Xiaohongshu's dots3-note
China's homegrown AI compute stack keeps closing the gap with the West — the latest sign: Huawei's Ascend platform ran Xiaohongshu's newest open model the day it shipped.
Huawei announced that its Ascend Atlas 800 A3 servers and Atlas 900 A3 SuperPoD supernodes now fully support dots3-note preview, the open-source multimodal model Xiaohongshu's Dots Studio released on August 14 — with "0-day" adaptation, meaning deployment support and performance tuning were finished the same day the weights went public.
The model itself is a big deal: 280 billion total parameters with 16 billion active under a mixture-of-experts architecture, handling text, vision and audio in a single model, with Xiaohongshu claiming it matches far larger models on reasoning, agent and multimodal tasks. We covered the launch — Xiaohongshu open-sources dots3-note, its IMO perfect-score LLM. What's new here is Huawei's move: the adaptation builds on the open-source vLLM Ascend inference engine, and Huawei says the full multimodal pipeline — visual encoder, audio feature extraction and the LLM backbone — runs intact on its hardware.
The engineering is mostly about squeezing performance out of Ascend's accelerators: fusing communication and computation steps in the MoE layers to cut inference latency, and native support for the model's MTP speculative decoding, which generates multiple candidate tokens per forward pass to speed up output under high concurrency. For developers, the practical upshot is that a deployment guide is already live in the vLLM Ascend documentation and the weights are on Hugging Face — no waiting for community ports.
Why it matters: 0-day support is the norm on Nvidia's stack, but it's still the exception on Ascend, China's main homegrown alternative. Huawei has been racing to court the open-source ecosystem — DeepSeek, Qwen and GLM all landed on Ascend before this — and dots3-note arriving on day one signals that Xiaohongshu's model is being treated as a first-class citizen of the domestic stack. The bigger story is the ecosystem itself: every major Chinese lab now expects its model to run on Huawei hardware at launch, and Huawei is increasingly competing on software maturity, not just chip specs.
What to watch: whether dots3-note's Ascend port shows up in production workloads — and how long the 0-day gap stays at zero for the next big open release.
Huawei's 0-day Ascend ports are becoming the default for China's top open models — does that change the calculus for labs choosing between domestic and Nvidia hardware? Tell us in the comments.
Sources: IT之家 · 网易 · vLLM Ascend docs · dots3-note preview (Hugging Face)
