Multimodal Qwen models released for image and audio
Alibaba unveiled three new multimodal models over the past week, extending the Qwen lineup beyond text into image and audio generation. The company released Qwen-Image-3.0, an image generation model that supports ultra-long inputs of up to 4.5k tokens—a roughly 4.5-fold increase over its predecessor—enabling one-step generation of knowledge diagrams, complex UIs, and multilingual posters across 12 languages and over 20 font styles. Alibaba also introduced two voice models: Qwen-Audio-3.0-Realtime, a conversational model with millisecond-level latency, tool calling support, and context-aware memory; and Qwen-Audio-3.0-TTS, a text-to-speech model supporting 16 languages and 20 Chinese dialects. The Plus version of the TTS model delivers 48kHz studio-grade output and topped the Speech Arena Leaderboard on Artificial Analysis. Alibaba said the models are aimed at applications in short films, comics, interactive education, and immersive entertainment.
Homegrown supernode runs 2.4-trillion-parameter inference
Alibaba Cloud said on July 23 that its Zhenwu M890 AI supernode has been adapted to run inference for the Qwen 3.8 foundation model, which has 2.4 trillion parameters, via the Bailian AI platform. According to Alibaba, the deployment marks the first time a domestically developed AI supernode has run inference for a model exceeding 2 trillion parameters. The Zhenwu M890 is Alibaba's latest in-house AI chip and supports precisions from FP32 to FP4. Its supernode architecture connects 64 chips through the proprietary ICN Switch 1.0 interconnect, delivering 800 GB/s bandwidth and 9 TB of aggregated memory—capacity suited to expert-parallel inference on large mixture-of-experts models. Alibaba said end-to-end optimization across chips, networking, and cloud software improved performance by as much as 1.5 times in agentic AI inference workloads. The Zhenwu M890 chip and Panjiu AI supernode server were named among flagship exhibits at the 2026 World Artificial Intelligence Conference earlier in the month.
Standalone Qwen Office product in internal testing
Alibaba is internally testing a product called Qwen Office, positioning it as a workplace AI offering separate from Tongyi Qianwen, according to reporting by TechNode citing Chinese outlet Lanjinger. The product is focused on intelligent collaboration for office workflows and may be released in the near term. Alibaba is recruiting solution architects, business development specialists, and AI product managers for the project. The company has been consolidating QoderWork, Wukong, and MuleRun into an enterprise productivity product under DingTalk CEO Chen Yusen, the report said.
Vertically integrated AI infrastructure strategy
The supernode milestone illustrates Alibaba's broader effort to compete in AI infrastructure through vertically integrated systems rather than individual chips alone. By combining proprietary AI chips, foundation models, and cloud services into a single stack, Alibaba is pursuing an alternative to Nvidia's ecosystem. Industry forecasts estimate China's domestic supernode market could reach 341.4 billion yuan (about $47.6 billion) by 2028, with Huawei, Tencent, Sugon, and Lenovo among companies investing in the segment. While maturity and developer adoption remain long-term challenges, the ability to run trillion-parameter models on a homegrown platform points to a more credible Chinese alternative for large-scale AI deployment.
Share this article
