Alibaba’s Qwen3.5‑Omni is closed
Alibaba launched Qwen3.5‑Omni, a multimodal model handling text, audio, images, and video — but unlike earlier releases, this one is proprietary, signaling a shift away from openness for high‑capability models. The move sharpens the open‑source vs proprietary AI debate around performance and accessibility. (theinformation.com)
Alibaba’s Qwen team published technical details for Qwen3.5‑Omni on March 30, 2026 in a Qwen blog post. (qwen.ai)) Alibaba is distributing Qwen3.5‑Omni through offline and real‑time API endpoints rather than releasing model weights or a public license for the Omni variant. (techinasia.com)) The release ships in three Instruct variants — Plus, Flash, and Light — and the Qwen team specifies support for a 256,000‑token context window. (qwen.ai)) Alibaba’s documentation and coverage say Omni can ingest more than ten hours of audio and over 400 seconds of 720p audio‑visual input at one frame per second, and it adds script‑level captioning with timestamps plus scene‑cut and speaker‑mapping outputs. (qwen.ai)) Alibaba and third‑party writeups list features such as voice‑cloning, “semantic interruption” for real‑time spoken interaction, and an emergent ability to generate code from spoken instructions paired with video. (qwen.ai)) Alibaba’s Qwen3.5 family previously included an open‑weight 397‑billion‑parameter release on Feb. 17, 2026 under an Apache‑2.0 style public license, while outlets noted the Omni rollout as a clear contrast — API‑only availability that commentators tied to Alibaba Cloud’s commercial AI push. (adcreate.com)) Smaller Qwen3.5 models and related tooling remain published on the project’s GitHub and on hubs like Hugging Face/ModelScope, while Alibaba points developers to Model Studio integrations and API compatibility layers (OpenAI/Anthropic‑style interfaces) for building with Omni. (github.com))
