Alibaba Cloud’s newest AI model, Qwen3-Omni-30B-A3B, has debuted with a splash. The Chinese company is touting it as “the first natively end-to-end omni-modal AI unifying text, image, audio & video in one model.” While Qwen3-Omni can accept prompts of text, image, audio and video, it only outputs text and audio. Alibaba Cloud has released the three versions of Qwen3-Omni so users can select based on their needs, choosing between general multimodal capabilities, deep reasoning or specialized audio understanding. Alibaba has also developed an AI chip called T-Head that performs comparably to Nvidia’s H20.

“At its core, Qwen3-Omni uses a Thinker-Talker architecture, where a ‘Thinker’ component handles reasoning and multimodal understanding while the ‘Talker’ generates natural speech in audio,” writes VentureBeat, explaining that “both rely on Mixture-of-Experts (MoE) designs to support high concurrency and fast inference.”

VentureBeat summarizes the three variants:

• The Instruct model is the most complete, combining both the Thinker and Talker components to handle audio, video, and text inputs and to generate both text and speech outputs.
• The Thinking model focuses on reasoning tasks and long chain-of-thought processing; it accepts the same multimodal inputs but limits output to text, making it more suitable for applications where detailed written responses are needed.
• The Captioner model is a fine-tuned variant built specifically for audio captioning, producing accurate, low-hallucination text descriptions of audio inputs.

Thinker is tasked with text generation while Talker focuses on outputting streaming speech tokens by receiving representations directly from Thinker, is how the Qwen team describes the process in a blog post, adding that “to achieve ultra-low-latency streaming, Talker autoregressively predicts a multi-codebook sequence.”

Streaming is central to Qwen3-Omni, says VentureBeat, detailing ingestion as “remaining under one real-time factor (RTF) even with multiple concurrent requests,” with “theoretical end-to-end first-packet latencies of 234 milliseconds for audio (0.234 seconds) and 547 milliseconds for video (0.547 seconds).”

VentureBeat says Qwen3-Omni may be Alibaba’s “most impressive model yet.” A demo allowing hands-on multimodal capabilities is hosted at Hugging Face Spaces, while the technical files for those who want to “download and build” are at Hugging Face Collections.

Meanwhile, Alibaba’s new T-Head chip was showcased in a broadcast on Chinese TV from the China Unicom Sanjiangyuan data center in Qinghai, writes Tom’s Hardware, noting that the T-Head’s accelerator “was directly compared with Nvidia’s H20 and A800, as well as Huawei’s Ascend 910B.”

Any degree of mastery over AI chip resources provides China with additional AI independence as it battles against embargoes on the top shelf offerings from the U.S.