Qwen's audio week: 3.8 Omni Flash sees and hears; LiveTranslate covers 60 languages

Alibaba released Qwen3.8-Omni-Flash and Qwen3.8-LiveTranslate across September 18–19, 2026. Omni Flash accepts text, images, audio and video with a 1M-token context and can reason over hour-long audiovisual input; LiveTranslate is a real-time speech-translation service covering 60 languages. Both are hosted services. Despite the Qwen name, there are no open weights to replace Qwen3-Omni-30B-A3B in a local stack.

Verdict: two strong real-time audio services, but neither is an open-weight successor to the local Qwen3-Omni model

The take

Qwen3.8-Omni-Flash outputs text, supports agentic planning and tool calls over audio and video, and accepts up to one hour of audiovisual input. Alibaba lists 113 input languages and dialects and claims an 89% lower video-input cost than Qwen3.5-Omni-Plus. Those are vendor claims; the concrete deployment fact is that the model is available through Model Studio, not as downloadable weights.

Qwen3.8-LiveTranslate supports 60 languages: 29 can produce both speech and text, while 31 produce text. Alibaba describes speaker separation, bilingual synchronized output and long-context disambiguation, and reports latency falling from 2.8 to 2.3 seconds under its LAAL measure. Again, this is a cloud API, not a local model card.

Our call: these releases matter for hosted live assistants and translation, but they do not move the local planner. `/models/qwen3-omni/` now names them explicitly as hosted successors so readers do not mistake the shared family name for a downloadable upgrade.

Where this fits

Models: Qwen3-Omni-30B-A3B-Instruct · Step-Audio 2 mini · MiniCPM-o 4.5 · Qwen3-ASR (1.7B / 0.6B)

Hardware: NVIDIA RTX 5090 · M5 Max MacBook Pro 64 GB

Sources

Next step

Try this in the planner→